Wolf Digest — Thursday, September 3, 2026

DisCo argues that what autonomous ML research agents lack is not a better harness but operational know-how, and makes that layer explicit as reusable verified skills distilled from code. Task-agnostic distillation over 1,000 widely used ML repositories yields the AREX-Skill Library: 5,000-plus verified skills across 20 areas and 178 capability families, complemented by task-oriented distillation at run time. Holding backbone, harness, and execution budget fixed, the skill-equipped agent gains 134.3% on MLE-bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet.

Benchmark distillation cuts how many agentic tasks you run; EarlyEval instead cuts cost inside each task by halting a rollout once its outcome is already predictable. A pair of LightGBM classifiers scores behavioral, textual, and reference-solution features at every step and stops execution when either crosses a calibrated confidence threshold. On SWE-bench Verified, TerminalBench, and Toolathlon this removes 13-26% of steps, up to 44.1% of input and 29.4% of output tokens, at 89-97% prediction accuracy, while resolve rates shift by only one to two points.

The Army moved its Enduring-High Energy Laser program out of prototyping with a $464.8 million other transaction agreement to AeroVironment, which the company characterises as the first production contract for directed-energy weapons in U.S. history. AeroVironment expects to deliver, in its phrasing, well into the dozens of LOCUST X3 systems over the next few years. The significance is less the dollar figure than the transition itself: counter-drone directed energy has spent a decade in prototype purgatory, and a production line changes the unit-cost and sustainment maths that has kept lasers out of fielded air defence.

Berlin concluded a month-long investigation by attributing the failed August drone attack on Leipzig/Halle Airport to Russia, alongside announced punitive measures. The explosive-laden drone landed near a Ukrainian cargo aircraft used to move munitions and did not detonate. Investigators cited signatures characteristic of Russian hybrid operations. For counter-drone programmes across Europe the relevant detail is the target set: a civilian logistics hub inside NATO territory, protected by airport security rather than air defence, which is precisely the seam that fielded counter-UAS systems have not been procured to cover.

DiffuSearch unifies the objectives of a two-stage driving planner rather than stitching together modules with mismatched goals. A guided diffusion model proposes a scene-consistent joint trajectory, using collision avoidance, drivable-area compliance, comfort, and progress as differentiable guidance during denoising; MCTS over a discretized action space then refines locally using those same terms as its reward. On nuPlan and interPlan reactive closed-loop benchmarks it reaches strong-to-state-of-the-art results, cutting collisions and improving comfort most in dense interactive scenes.

Meta's fourth Muse Spark release in five months targets long-horizon agentic work rather than raw capability. The model is trained across diverse harnesses to generalise between agentic environments, and trained to collaborate: asking clarifying questions on ambiguous prompts, invoking help when stuck, and confirming before consequential actions. Meta engineers report roughly 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2 on comparison coding tasks. It ships today in Muse Code and the Meta Model API, with max reasoning gated behind unfinished safety testing and an open-weights release signalled on the roadmap.

Cliff sharpens RLVR credit assignment without a trained process reward model or a teacher that shares the student's reasoning style. An off-the-shelf LLM only has to locate the first mistake in a rollout, splitting it into a valid prefix and an invalid suffix; everything after the break carries little information because it is conditioned on a broken state. That split becomes token-level advantages, positive before the cliff and negative after. Across 12 settings it beats on-policy distillation by 15% and GRPO by 7%, holding up even with weak teachers.

A decade of U.S. policy has aimed at rebuilding domestic leading-edge fabrication capacity to cut geographic concentration risk, and CSET's report makes the point that capacity is not the binding constraint. Fabs cannot run at scale without a steady supply of technicians with cleanroom-specific skills and high-reliability process discipline, which takes years to build and does not transfer from adjacent manufacturing. It pairs closely with RAND's space-workforce finding published the same day: two separate strategic-industry analyses landing on personnel rather than capital as the limiting factor.

Clinical telerobots carry a single stereo endoscope, so multi-viewpoint surgical data barely exists and sparse-view 3D reconstruction methods have never been tested on real endoscopic imagery. MV-dVRK is an ex-vivo dataset with exposure-synchronized stereo viewpoints, dense SfM geometry validated against an industrial 3D scanner, and ground-truth poses. Two endoscopes give multi-stereo the best coverage; with a third view, optimization-based multi-view methods cover 67% of ground-truth surface within 1 mm versus 43% for feed-forward foundation models. Ten dynamic sequences with tissue deformation are included.

SolarWM releases an end-to-end open stack for interactive video world models, targeting the mismatch between heterogeneous video corpora and backbone-specific training code. Its data engine normalizes 1.43 million clips from 10 datasets into a frame-aligned contract carrying observations, metric camera geometry, captions, quality metadata, and provenance, decoupled from mixture construction. Four models spanning 5B to 33B are built on Wan2.2, LTX-2.5, and MiniMax-H3 under one three-stage recipe of bidirectional adaptation, teacher-forced autoregressive init, and distribution matching distillation, yielding causal models that sustain real-time rollouts of minutes to hours despite 5-second training sequences.

Routing a training sample to the teacher whose domain label matches it assumes expertise that only holds on average. MT-SDPO instead picks the teacher per sample by verification: a frozen teacher may supervise only if its own answer passes a verifier, rollouts are additionally anchored by a correct sibling from their own group, and anchor plus verified feedback are merged into a privileged context read by an EMA self-teacher the student never sees, so one policy ships. Across five students from three families it lifts the weakest domain of Qwen3-8B by 14.79 points and narrows the domain gap by 74.7%.

A quiet GSA website update replaced USAi's no-cost model with a platform fee plus passthrough usage costs. The platform was the main on-ramp for agencies experimenting with commercial models without running their own procurement, and free access was what made adoption legible in usage statistics. Introducing a fee moves the decision back to individual agency budgets, which is likely to slow the uptake numbers the administration has been citing and to advantage agencies that already have direct vendor agreements. GSA did not respond to requests for comment on the change.

Russian strikes on Ukraine's terrestrial communications deepened Kyiv's dependence on a single commercial satellite constellation for command and control, drone operations and artillery coordination, with no European alternative available if access were withdrawn. The essay generalises from that to state-business relations across critical digital infrastructure, where operational capability sits with private firms whose commercial decisions have become defence-relevant, and where procurement instruments designed for equipment do not give governments the assurances they now need for services.

The government's brief states that the United States has a strong interest in continuing to develop a competitive AI industry that sets the global standard for how AI is used, and aligns with OpenAI on the training-data question. Whatever the eventual judicial outcome, an executive-branch filing on the defendant's side changes the litigation landscape for every pending training-data case, and it removes a source of uncertainty that has shaped data-sourcing decisions and licensing spend across the industry for two years.

The committee voted to subpoena Oracle Health executives after a hearing on the Department of Veterans Affairs electronic health record modernisation programme, following a $17 billion increase in the contract ceiling that drew bipartisan objection. The programme is the largest civilian health IT modernisation in the federal portfolio and increasingly the venue where questions about vendor-supplied clinical AI and decision support get litigated, so the oversight posture here sets expectations for how Congress will treat similar contracts.

Borrowing de Finetti's theorem, this work measures LLM forecast coherence without ever needing outcome labels: elicit probabilities over related events derived from stock-return data, then solve a linear program for the maximum guaranteed Dutch-book profit an arbitrageur could extract. Models are substantially incoherent, and incoherence grows as the logical relationships among events get richer. Most strikingly, irrelevant contextual detail in the prompt can inflate the arbitrage bound by an order of magnitude. The label-free construction makes this usable on unresolved or unobservable events.

GDB-Reward turns a grab-bag of graphic-design evaluation metrics -- typography, layout, color, spatial placement -- into a single scalar RL reward used to optimize prompts for a completely frozen text-to-image generator, sidestepping diffusion fine-tuning entirely. Because RL only needs a ranking signal rather than gradients, non-differentiable design checkers work directly as objectives. Reported gains span perceptual quality, text-rendering fidelity, and spatial accuracy against the design spec. The broader claim is that evaluation harnesses in any domain lacking differentiable supervision can be recycled as training signal.

SCX Router scores the suitability of each candidate inference endpoint without autoregressive generation, using a 0.6B GLiClass-style checkpoint that pairs a Qwen3 decoder with a shallow bidirectional scorer. Its decoder-KV path keeps a text-only KV cache across a session, encodes only new turns, and evaluates candidate-label tokens transiently so they never pollute the persistent cache; the same head also predicts task type, difficulty, reasoning mode, and expected output length. Training data comes from an ontology of 23 families, 115 task types, and 345 routable subtypes expanded into 150,000 verifier-scored tasks.

Using the 8,627-question clinician-reviewed MedMisBench reasoning subset, this study injects two misleading cues, fabricated evidence and a bare assertion, into medical QA and tracks susceptibility, disclosure, and monitorability across three reasoning models. All three adopt the asserted answer 10 to 27 points more often than the fabricated-evidence answer, yet the assertion is the cue least likely to be disclosed: cues appear in 81-98% of traces but only 7-90% of responses. Resampling shows evidence corrupts reasoning early and cumulatively while assertions redirect the conclusion late; an LLM monitor reading open traces catches 78% of corrupted decisions at 5% FPR versus at most 32% from responses alone.

Artificial Analysis places Muse Spark 1.3 (xhigh) at 61 on its Intelligence Index, up 4 points from 1.2 in August and 8 from 1.1 in July, tying GPT-5.6 Sol (max) and Grok 4.6 (high). The partner-only max variant reaches 62. Gains concentrate in agentic evaluations: Tau3-Bench Banking climbs 35 to 47 percent for xhigh and 52 percent for max, the top score recorded on that evaluation; Terminal-Bench 2.1 rises 80 to 85 percent; GDPval-AA v2 moves from 1615 to 1709 Elo. CritPt gained 8 points and GPQA Diamond 4. Two regressions: AA-LCR fell 4 points to 79 percent for both variants, and AA-Omniscience accuracy dropped 3 points for xhigh, traced to a higher abstention rate that also cut its hallucination rate. At $0.55 per Index task, no model scoring 59 or above costs less.

CMS chief information officer Patrick Newbold said the agency has moved past counting how many staff use AI tools and is now measuring what the tools contribute to outcomes. It is a small administrative signal with a real consequence: seat counts and prompt volumes are the metrics most federal AI reporting still runs on, and they are trivially gameable. An agency of CMS's size adopting outcome measures gives other departments an internal precedent to point at, and it raises the evidentiary bar vendors have to clear at renewal.

World models for web agents are usually trained by supervised next-state prediction, an objective misaligned with how they are actually used -- a ranker needs predicted states that are discriminative between candidate actions, not accurate HTML or AXTree reconstructions. Predicted-state matching trains the model to distinguish the true resulting state from states reached by alternative actions, using a branching dataset built from WebArena Go-Browse trajectories where each decision point has multiple actions and outcomes. It beats supervised next-state models on held-out matching, improves ranking on WebPRMBench over action-only PRMs, and raises end-to-end success on WebArena-Lite.

HiPoly builds polymer structure into the architecture rather than flattening it into a small-molecule representation. A three-level hierarchical graph over the G2RINS representation encodes stochastic inter-monomer connectivity, composition, and molecular weight, and the same representation drives property prediction, generative design, and MD-based validation end to end from experimental formulation data. It reports state-of-the-art thermophysical prediction on multi-component systems, with ablations showing each hierarchy level contributes independently. The generative path was used to propose and independently validate PFAS-free replacements for fluorinated polymers hitting target surface-energy specs.

DroneCATS puts an MLLM directly in a drone control loop with the entire action space declared in the prompt, no fine-tuning and no function-calling schema, and treats the model as the swappable independent variable across approaching, tracking, searching, and multi-drone commanding. Scaling the roster down to 2B parameters isolates what breaks first, and it is not flight control: small open models often reach the success radius more reliably than frontier models but lose episodes by declaring arrival early or never. Multi-drone commanding widens the gap, with small models copying one coordinate across distinct views.

Independent evaluation puts Gemini 3.8 Flash at 59 on the Artificial Analysis Intelligence Index with high reasoning, 3 points above 3.7 Flash and level with sub-maximum efforts of GPT-5.6 Sol and Grok 4.6. Medium reasoning scores 57 and low scores 52, the latter matching Gemini 3.6 Flash at high reasoning for 30 percent less cost and roughly a third the time. The gain is concentrated in agentic evaluations, led by a 12-point jump on tau3-Banking to 45 percent. Cost per task rises about 40 percent to $0.58 despite unchanged per-token pricing, driven by output tokens climbing 30 percent to 48k per task and more turns on agentic runs. Output speed stays near 300 tokens per second, but time per task rises from 2.2 to 2.5 minutes.

Loom targets industrial root cause analysis, where monolithic agents hit context limits and latency budgets while classical weak supervision cannot handle free-text hypotheses. Modular diagnostic templates populated with episode-specific entities, times, and metrics emit open-form hypotheses that Loom projects into embedding space and reconciles with iterative centroid-based reweighting, the resulting weights grounding one lightweight synthesis call. On OpenRCA it matches a state-of-the-art autonomous agent on Bank and Market-2 and trails on Market-1 and Telecom, at one LLM call per incident, roughly 26 times faster and 33 times with an 8B synthesizer.

Persistent agent memory is shown to be part of an agent's effective authorization policy, not just a performance cache. The failure mode named here, endogenous authorization laundering, needs no attacker: incremental memory writes record permissions the underlying event history never granted, and provenance is lost. EAL-Bench evaluates five LLMs as memory writers and two as executors across procurement, cybersecurity, and finance; writers manufacture false authority for up to 50.2% of unauthorized requests, and executors act on that false authority in 98.6% of trials. Requiring source-event backing and bounded event sourcing both curb laundering while rejecting more legitimate actions.

DKL injects corpus knowledge into an instruct model without the usual tax. Rather than continued pre-training on the instruct checkpoint -- which erodes instruction following and forces a costly IFT pass that may lack a suitable corpus -- it runs extended pre-training on the corresponding base model and merges the knowledge-infused weights back into the instruct model. This avoids the massive synthetic QA generation that RAFT and PA-RAG require. On retrieval-failure cases, where RAG breaks down, accuracy rises from 54.17 to 79.26 while instruction-following capability is preserved.

Influence functions estimate the effect of infinitesimally reweighting a training example, and under weight-based interventions the examples they surface often barely beat random selection. This paper separates selection from intervention: keep the influence-selected examples but rewrite their responses with behavior-aligned or behavior-opposed supervision while holding instructions fixed. Across four open-weight LLMs, tested primarily on epistemic abstention and extended to safety refusal, rewriting yields stronger, more persistent, bidirectional behavioral shifts where reweighting the identical examples is weak and inconsistent, and effects stay concentrated on target behaviors.

PaperCompiler inserts an explicit compilation step between paper and repository: instead of free-form plans that downstream coding agents can quietly compress or reinterpret, it emits repository-level specifications encoding non-degradation requirements, file ownership, cross-file dependencies, and provenance that distinguishes paper-supported, inferred, delegated, and unresolved facts. Generation then runs under those constraints while leaving local engineering choices free. On Paper2CodeBench it lifts reference-based fidelity 13.8% relative (3.64 to 4.15) and halves high-severity evaluator critiques from 13.2% to 6.1%.

RAND's perspective treats training on a stronger model's outputs as a strategic rather than merely commercial problem, arguing that distillation compresses the window during which a frontier lead translates into national advantage. The technical countermeasures available to labs, chiefly raising the cost of extraction or degrading output quality for suspected distillation traffic, trade directly against product quality and are individually weak. The authors argue policy instruments need to augment them rather than wait for a technical fix, which puts terms-of-service enforcement, access tiering and export-adjacent controls into scope for a behaviour that currently sits almost entirely outside regulation.

Shield AI's post pulls apart the word simulation, which currently covers desktop trainers, fixed-base and full-flight simulators, mission rehearsal and test environments for systems that do not exist yet, and argues these are built to different fidelity requirements and should not be procured as one category. The relevance to autonomy is direct: the same visual-simulation stack that trains aircrew is where autonomous behaviours get validated before flight, and the fidelity gap that a human pilot tolerates is exactly the distribution shift that breaks a learned policy.

Michael Hunzeker and Yuster Yu revisit their 2025 argument that Taiwan's readiness problem is cultural rather than material, spanning both the officer corps and civilian society, and assess what has and has not shifted in the year since. The AI-adjacent thread is mobilisation infrastructure: societal resilience arguments increasingly turn on communications, information integrity and civilian-facing command tooling, which is where the technical and the cultural questions meet.

SafeEvolve treats agent safety as a property of the harness and the policy jointly, closing a loop where completed on-policy trajectories become safety evidence. On the harness side, trajectory-level evidence is compiled into bounded component-level edits to the safety prompt and hierarchical skills, producing auditable and reversible artifacts rather than opaque prompt drift. On the policy side, harness-use SFT teaches the model to invoke those artifacts, then harness-augmented RL with verifier-decomposed rewards shapes multi-step behavior. On agentic safety benchmarks it improves the safety-utility frontier, with a 3x attack-success-rate reduction reported for Qwen3.5-4B.

Streaming video work has attacked token pruning, merging, quantization, and frame retrieval but left model depth alone, so every incoming frame still pays a full-depth MLLM prefill and grows the KV cache in proportion. ShallowStream runs only the shallow layers continuously, reusing their KV cache as an always-on retrieval index, then uses shallow-layer attention scores plus a diversity-aware selection rule to pick evidence frames for full-depth answering. It matches the strongest streaming baselines while cutting per-frame prefill and 10-second end-to-end latency by up to roughly 52%.

Sampled-token on-policy distillation reliably lifts pass@1 while pass@k flatlines -- the student inherits the teacher's mode but not its diversity. The authors diagnose this with First-Order Local Entropy Influence, a signed proxy that splits each update's entropy effect into the teacher-student log-prob gap and the student's local probability structure, tying entropy collapse to negative-influence token positions. IDA-OPD keeps entropy-expanding updates and replaces contracting ones with divergence-adaptive advantage shrinkage, using only sampled-token teacher log-probs. It restores pass@k without full-vocabulary forward-KL cost while holding pass@1.

Leech-lattice VQ posts the best reported 2-bit weight quality, but no one had implemented the multi-shell decoder the rate actually needs. This work supplies one -- an offline expansion of the 301-class codebook into GPU layouts plus a fused dequantize-and-GEMV kernel with no warp divergence -- and measures batch-1 decode serving cost. It separates in-VRAM rate from on-disk rate: binary bit planes beat one-hot masks at 4.80 bits per weight (2.15x FP16). A deployed QTIP trellis kernel reads 2.40x fewer bytes and runs 2.27x faster, with end-to-end gains of 1.11x, 1.29x, and 1.41x at 4B, 8B, and 14B.

The received view is that naturally occurring user feedback is too noisy to train on. This paper argues the noise is largely in the measurement: using synthetic data with definitive ground truth alongside naturalistic data, revisions produced with feedback resolve the targeted issue at substantially higher rates than revisions without it. The sharper finding is the failure mode in evaluation itself. When a model fixes an issue specifically because of feedback, LLM judges frequently fail to recognize the corrected response and prefer the inferior baseline, meaning judge-based pipelines will systematically under-credit feedback-driven improvements.

The All-domain Anomaly Resolution Office moved to issue a sole-source contract for a commercial data subscription to a New Mexico nonprofit holding what is reported to be North America's largest private archive of unidentified anomalous phenomena records, including case files with no government counterpart. Setting the subject matter aside, the mechanism is the notable part: a defence analysis office buying a private historical corpus as a subscription is the same acquisition pattern now being used to feed government analytic pipelines generally, and sole-source justification for a unique dataset is the argument that will recur.

Intrusion detection for connected vehicles has focused on infrastructure-side defense or V2V basic safety message misbehavior, leaving onboard verification of Signal Phase and Timing messages unaddressed. SPADE is a labelled multi-modal simulation dataset built in Eclipse MOSAIC with runtime attack injection at the SAE J2735 application layer, covering six attack classes plus benign traffic. Four intersection geometries, six operating conditions, and five random seeds give 180 base runs and roughly 1,890,000 labelled timestep records, 270,000 per class, each fusing SPaT fields, onboard camera confidence, and cooperative peer data across 40 features.

TRACE targets post-incident auditability of learned robot policies by splitting decision-making into four documented layers: evidence-grounded semantic perception, probabilistic belief reasoning over causal graphs, constraint-aware action synthesis with counterfactual records, and execution verification. It is model-agnostic, wrapping CNN or transformer perception without changing them. On 500 simulated warehouse navigation decision cycles it reports 98.6% evidence traceability, 99.0% temporal continuity and 98.1% decision reconstructability. The authors argue attribution methods such as LIME give feature weights but not the artifact structure regulators would need to reconstruct a decision.

McGrath, Goodfire co-founder and former DeepMind researcher, argues for interpretability as a route to new scientific knowledge rather than post-hoc explanation of model outputs. The conversation runs from learned modularity in AlphaZero through neural geometry: concept manifolds, reusable computation recurring inside Llama, and why activation steering succeeds and fails in the patterns it does. The framing question is whether internal representations converge on structures that exist in the world, which is the claim that would make extracted features scientifically informative rather than merely diagnostic.

RIG-BENCH tests whether unified generative models can infer a latent rule from visual input and then render a solution that satisfies it, rather than merely aligning surface events. Its 2,000 curated samples span concept-based, transformation-based, pattern-and-structure, and scenario-based tasks. Across current unified models and image/video generators the authors find a consistent reasoning-generation gap: outputs are locally plausible but globally inconsistent with the inferred rule. The benchmark is positioned as a diagnostic for world-simulator-style systems where logical grounding, not fidelity, is the bottleneck.

LoRA quality is unusually sensitive to initialization, and prior schemes seed the factors from principal components of pretrained weights, activations, or gradients -- none of which reference the full-rank model's actual training dynamics. TaRA instead picks an initialization so that the gradient induced through the low-rank factors approximates the gradient of the corresponding full-rank weight matrix, derived in closed form and adding negligible overhead. It outperforms prior state-of-the-art initializers across a range of difficult fine-tuning tasks, a cheap drop-in change to a standard PEFT recipe.

The paper names "linguistic illegibility": the gap between an LLM's externalized or mechanistically probed language artifacts and the activation-space computation actually driving behavior. If that gap is always possible in principle, then any security control resting on the model's linguistic self-report, chain-of-thought monitoring, constitutional self-critique, activation probing for linguistically defined features, can never be fully sound. The proposed alternative is sandboxing whose guarantees never read the model's linguistic state at all, specifically taint tracking that declares a priori which system state must never be influenced by model-produced data.

Probing five open-weight models on matched valid/invalid premise-claim pairs separates three things usually conflated: representing logical validity, expressing it behaviorally, and using it causally. Behavioral accuracy sits near chance, yet validity is nearly perfectly decodable from hidden states and stays strongly decodable under held-out templates, semantic domains, and inference families, including on examples the model answers wrongly. Exhaustive leave-one-out tests bound that generalization, and steering along probe-derived validity directions produces only weak, nonspecific effects relative to random controls, a useful caution against reading causal structure off linear probe accuracy.

This work makes Hessian-style sensitivity analysis tractable at billion-parameter scale with a Kronecker-factored Fisher approximation that captures cross-layer interactions without materializing the full matrix. The analysis surfaces a consistent pattern across model families: value projection layers are the most sensitive and most strongly cross-correlated, while other components behave in architecture-specific ways. Experiments on quantization, sparsification, inter-layer corruption, and post-corruption fine-tuning show the approximation tracks both degradation and recovery, giving a principled signal for mixed-precision allocation, layer-wise sparsity, and adaptive low-rank decomposition.

Trajectory-level GRPO smears one terminal reward across an entire free-form reasoning trace, so credit never lands on the specific visual deduction that decided the answer. Credit-addressable reasoning makes the inference-time semantic units the same units learning compares: Code-CoT keeps the diagram and expresses visual relations as line-addressable executable code organized into typed events, while CE-GRPO cuts event boundaries using structural priors and type-normalized entropy, samples continuations from shared prefixes, and converts outcome differences into local advantages. Average accuracy across nine geometry benchmarks reaches 76.04, beating Qwen3-VL-8B by 8.09 points and trajectory-level GRPO by 3.43.

The post builds the argument from the ground up: a single forward pass reasons broadly but shallowly, chain-of-thought buys depth by serialising computation into tokens, and that serialisation is incidentally the only reason external monitoring works at all. Neuralese refers to models passing continuous latent state between steps rather than discrete tokens, which recovers the depth without the legible trace. The timing matters, since it supplies the shared vocabulary for the Astra recurrence debate that broke the same day.

CORAL puts an LLM agent inside the continual tuning loop of a live production recommender rather than in offline model development. Each cycle the agent reads operating signals, reasons over a memory of prior decisions and their measured effects, and calls tools including a numerical optimizer that holds changes inside a fixed operating budget. The setup is framed as partially observed, non-stationary constrained optimization where the policy improves in context with no parameter updates. A/B tests on two large social platforms show engagement gains at flat serving cost on one and cost reductions at flat engagement on the other.

CRISP attacks two weak points in dynamic sparse-attention routing for long-context prefill. It swaps Jensen-Shannon-divergence head routing for C_struct, a structural score that reads mass at Vertical-Slash-compatible positions directly off the proxy map, reproducing JSD's routing decisions while dropping the pooled matmul and the KL computation. It also replaces cumulative coverage thresholds, which the authors show accumulate O(n) background noise at long context, with a sink-aware threshold anchored to the noise floor. Across InfiniteBench, RULER and LongBench on two model families it is the strongest sparse method and matches or beats dense attention on retrieval-heavy tasks.

Palaestra Research is releasing MonitoringBenchHonest, a benign-trajectory companion to the attack dataset MonitoringBench. The gap it closes is the one that makes monitor evaluations misleading: a detector scored only on attack trajectories has no measurable false-positive rate, and a monitor that flags everything looks excellent. Pairing honest runs against the existing attacks turns monitor comparison into a proper operating-characteristic question, which is what deployment decisions actually need.

A new convoy of Russian armoured vehicles arriving via Lomé marks the fifth year of the Africa Corps deployment in Mali, and the analysis asks what return the investment is meant to produce. Wagner entered in December 2021 pursuing the resources-for-security arrangement it had reached in the Central African Republic; the piece traces how that model has and has not transferred, and what the mission's cost structure implies about how long it can be sustained.

BLUEPRINT factorizes multi-turn jailbreaking into 18 theory-grounded social-influence factors and separates them from WORLDVIEWSIM, a cross-turn situational context module, then uses MCTS to optimize turn-level combinations over four-turn trajectories. Against six frontier models it reaches near-ceiling attack success with the fewest average queries at 2.46, and the search traces expose model-specific vulnerability profiles. All targets share one recovery pathway out of hard refusal: reframing toward concrete, executable tasks. Ablations rank actionability first and gain framing unexpectedly high, while some legitimacy appeals reduce success.

Counter-GEO-Bench treats generative engine optimization as an attack surface: adversaries publish ordinary-looking optimized pages that generative search retrieves and synthesizes into distorted answers. It pairs 247 quality-gated human-verified queries with both information-preserving and information-distorting GEO rewrites and scores defenses on attack success rate, false positives, and answer quality across three victim LLMs. Granite Guardian, Llama Guard 3, and NeMo self-check fact-checking cut ASR by at most 5.7% relative, since safety taxonomies flag policy violations and this content reads as fluent information. A purpose-built baseline, C-GEO Guard, reaches 47.6% relative ASR reduction at near-zero utility cost.

ExecRetrieval plants execution-verified counterfactuals directly in the retrieval pool: 939 Python tasks, each with a canonical implementation plus up to four single-edit mutants that are near-identical lexically but fail tests. Across 23 dense embedding configurations plus BM25, the strongest hosted system hits exec@10 = 1.00 but exec@1 = 0.331, and its rank-1 miss is a paired buggy variant 91.5-99.4% of the time; the canonical ranks below at least one paired distractor in 67-78% of queries. The result isolates a functional-correctness blind spot that topical benchmarks cannot surface, with direct consequences for RAG-based coding agents.

Training LoRA factors independently ignores the geometry of the low-rank weight change they induce. LoRA-TSD treats each step as a tangent vector on the fixed-rank matrix manifold, takes Muon's spectral-norm steepest-descent step inside that tangent space, and retracts back to the factors through a LoRA-native retraction that is up to 2.8 times cheaper than truncated-SVD retraction. The Frobenius variant of the surrogate is shown to recover LoRA-Pro, and the paper gives the first global convergence guarantees for both. It beats all competing LoRA optimizers across six commonsense and NLI benchmarks on Llama-3.2-1B, Llama-3.1-8B, and Qwen3-32B, robustly across adapter ranks.

A dynamical model of AI-assisted AI R and D yields a recursive reproduction number, R_AI, comparing the strength of capability feedback against the rate at which further research gets harder. Above 1 improvements compound across development cycles; below 1 they damp. The threshold is a property of the feedback loop's structure, not of any capability level, so a system can already be self-amplifying before acceleration is visible, and fast progress can occur with no self-amplification at all. Higher baseline productivity speeds progress without changing the regime, cycle duration sets the amplification timescale, and shared improvements can make a multi-actor ecosystem self-amplifying even when no single lab is.

Semantic entropy and token log-probability uncertainty fail in complementary ways for black-box hallucination detection: the former is uninformative when all sampled responses land in one semantic cluster, the latter misses confidently wrong answers. This work aggregates token-level signals across samples via a TopK method and adds two supervised combiners -- Gated, which routes single-cluster cases to a token-feature classifier, and Stacked, which learns jointly over semantic and token features. Across seven benchmarks and four models, Stacked wins in nearly half of settings while TopK and CoCoA stay competitive with no labels.

VibeVoice-ASR-Streaming pushes unified speaker-attributed ASR from offline to streaming, one of the first LLM-based end-to-end systems to do so. It interleaves fixed-size audio chunks, a small lookahead window, and previously emitted text, so who-said-what is produced incrementally without a separate diarization stage, which is the blocker for real-time voice assistants and agents. The 7B model records the lowest average WER/CER across five evaluation sets and best or tied-best speaker attribution on 12 of 13 settings. Both 1.5B and 7B weights plus inference code are released.

Using power-law kernel regression in the one-pass regime, this analysis derives critical learning rates for SGD, Polyak, and Nesterov momentum as functions of batch size B, momentum rho, and capacity exponent beta, then scaling laws for the full risk trajectory from early transient through power-law decay to noise floor. Minimizing final risk under a fixed data budget yields a three-regime batch-size phase diagram: Polyak enlarges the critical batch size, buying parallelism without losing the small-batch data-scaling exponent, while Nesterov's look-ahead suppresses noise accumulation and wins on data efficiency at large batch. Numerics confirm the predicted stability boundaries.

MultiGhostBench pushes LLM authorship attribution to book length: 928 machine-written books from five recent models across six languages and three scripts, averaging about 59K words each, with explicit domain, author, and language shift splits. No attribution method wins across settings and all degrade under distribution shift. Transformer-based detectors carry generator-specific signal across languages, though transfer quality depends on the language pair, while statistical and fingerprint methods stay tied to the language they were fit on. Dataset and code are public.

UTP-Bench adds stochasticity to LLM travel planning, which TravelPlanner and TripCraft treat as deterministic constraint satisfaction. It covers 504 Indian cities with attractions, lodging, and multi-modal transit, layered with empirical delay distributions and crowd-density patterns, and scores itineraries on three new robustness metrics (buffer adequacy, crowd-aware timing, transport delay absorption). GPT-5, Qwen3, Mistral, and Phi-4 all trail human-authored plans, with the largest deficits in temporal buffering -- models schedule tightly and have no slack when transit slips.

Hyper-Connections run n parallel residual streams mixed by a learned n by n matrix, and leaving it unconstrained lets rescaling compound across depth. The manifold-constrained fix restricted the matrix to doubly stochastic matrices, capping amplification at one but leaving contraction unbounded. The authors prove that within that set the mixing can only shrink norms by shrinking inter-stream differences while leaving their mean fixed, so stream diversity is spent with depth. Their oHC instead constrains the matrix to SO(n), and for the four-stream case parameterizes it in closed form with a pair of unit quaternions, adding no parameters and replacing iterative projection with a fixed pattern of signed additions.

Anthropic publishes the system prompts for its consumer applications along with the historic diffs, and Willison's read of the current version finds the copyright section carrying disproportionate weight, with song lyrics singled out repeatedly rather than handled by a general rule. The observation worth keeping is structural: published prompts plus published diffs turn deployment-layer policy into something external researchers can track over time, which is a form of transparency almost no other lab offers and which reveals where legal exposure, rather than capability, is shaping model behaviour.

SonicCaps attacks the one-caption-per-clip bottleneck in audio-language pretraining with roughly 15M captions over about 700k clips, around 24 per audio, produced by Qwen3-Omni conditioned jointly on audio and text and structured into main descriptions, verbosity and style rephrasings, and semantic tags. Human raters score it above existing captioning corpora on descriptiveness and precision. CLAP models trained with multi-caption sampling improve audio retrieval and zero-shot classification on both public and commercial benchmarks; dataset and two CLAP checkpoints are released.

The method takes a one-line natural-language description of a target behaviour and produces on-policy examples of it using only logit access, with the accompanying LogitTilt technique released as tooling alongside code and transcripts. The practical appeal for evaluation work is the input cost: writing a sentence is cheap enough that behaviours can be enumerated speculatively rather than hand-curated, which is the bottleneck that keeps behavioural test suites small and stale.

CoMerge recasts multi-task model merging as preference optimization, using the failure modes of naive merging such as task arithmetic as hard negatives to build preference pairs with no external annotation. Only lightweight tensor-wise merging coefficients are optimized, 1,445 scalars in the reported setup. It reaches 0.9968 average normalized performance on MergeBench, ahead of all data-free and data-driven baselines, and on Llama-3.1-8B-Instruct it improves conflict-sensitive behaviors like instruction following and safety while staying competitive with full-parameter fine-tuning.

AICOME proposes treating LLM-derived respondent-level measures as inputs to contextual models rather than as prediction targets, decomposing each measure into a group-level aggregate and an individual deviation so between- and within-group associations can both be estimated. Validation uses the 2022 China Family Panel Studies with occupations as the grouping structure and job-related survey items as benchmarks, covering computer use, foreign-language use, weekly hours, and management responsibility. Weekly hours validates most cleanly, with AI-derived measures reproducing the large negative between- and within-occupation associations with satisfaction seen in the survey data.

SimLoss adds a reference-free training objective that pushes a VLM's projected hidden state toward a frozen image embedding under InfoNCE, injecting dense visual supervision before any token is decoded. This removes the need for human-written fine-grained captions or pseudo-captions from generate-decompose-verify-rewrite pipelines. Two variants are given: a fully differentiable fine-tune through a local embedding model, and a GRPO version treating that model as a black-box reward. The differentiable variant gets the highest precision of all baselines and nearly matches multi-stage F1 while running about 20 times faster at single-pass inference.

CivBench wires Civilization VI to language model agents over MCP, exposing 76 tools plus a narration layer that turns visual game state into structured text; a single episode runs past 300 turns and thousands of tool calls under partial observability. With only 23 admissible runs across four model families the authors explicitly decline to rank models, and instead propose two interface-level metrics: Proactive Monitoring Rate and RAG@10, which checks whether stated planning commitments execute within ten turns. Agents systematically under-query available state, checking victory progress every 30 to 75 turns against a 20-turn playbook target, and missed the warning window entirely in 7 of 20 detectable defeats.

GitHub's argument against optimising raw token counts is that a terse tool response frequently costs more end to end, because it forces extra calls or a retry that a fuller response would have avoided. The team frames the objective as supplying the right context to advance a task rather than the least context, and reports efficiency work aimed at turns and completed tasks instead of per-interaction tokens. It is a useful corrective for anyone benchmarking agent cost, since token-per-call is the metric most harnesses expose and the one that rewards exactly the wrong behaviour.

Choosing a retriever for production RAG needs repeatable comparative judgments, which get expensive every time a new candidate arrives. This work applies classic pooling to LLM judging: the model judges the union of documents returned by current candidates, and the pool grows incrementally as only newly contributed documents are judged, with all judgments reused across systems. Validated on four benchmarks with 11 dense, sparse and hybrid systems, pooled rankings correlate strongly with gold qrels and preserve 97% of pairwise orderings under bootstrap uncertainty. Deployed across 62 configurations for financial news QA, document overlap gave 65-80% judgment reuse and up to 4.9x lower evaluation cost.

A modular pipeline built with Boston Public Library turns historical newspaper scans into structured pretraining-grade text, deliberately kept frugal enough to run on workstation hardware and interpretable at every step. Each scan is segmented into type-agnostic crops, OCR'd per crop, then passed through text analysis, type classification, reading-order detection, NER, subject classification, language ID, and embedding generation. The released open dataset covers 1,473,635 public domain scans published 1795-1930, producing 83.1 million crops and 16.3 billion o200k_base tokens, alongside the small task models trained for each stage.

Animal motion modeling has been starved of data because mocap rigs do not generalize past humans. Kirin reconstructs 3D motion from in-the-wild quadruped video at scale and pairs it with captions to build AiM3D, described as the first large-scale aligned video-text-motion dataset for quadrupeds. On top of it the authors train a generator conditioned jointly on text and image, then pipe outputs through an off-the-shelf image-to-3D model to auto-rig and animate meshes. The result is an end-to-end path from a prompt to a render-ready animated animal, without species-specific capture.

AffectDelta reframes emotion-driven image editing as a transition between eight-dimensional emotion distributions rather than a jump to a single categorical label, so the edit signal encodes which affective components should rise, fall, or hold. A frozen emotion-distribution predictor estimates the source state and the signed source-to-target delta conditions a transition encoder feeding a source-aware diffusion backbone. Training uses AffectPair-249K, 248,841 source-target pairs covering cross-category and within-category transitions. Against six baselines it improves affective alignment while preserving scene composition, with ablations supporting the delta conditioning.

Arc Science Fellow Jingtian Zhou has built both halves of the problem: wet-lab assays that measure how DNA physically folds inside the nuclei of individual cells, and the algorithms that turn that measurement into something interpretable. The reason cell-type identity from identical genetic code is a modelling problem at all is that the signal lives in three-dimensional contact structure, and single-cell contact maps are sparse enough that the analysis method determines what is recoverable as much as the assay does.

How much can predetermined stepsize schedules accelerate plain gradient descent on smooth convex problems? This paper pushes past the classical Nemirovsky-Yudin first-order oracle bound with an Omega(n^-1.6342) non-anytime lower bound and an Omega(n^-1.2408) anytime lower bound, improving prior results of Omega(n^-1.932) and Omega(n^-4/3). Combined with the known silver-schedule rate, the anytime bound establishes a strict separation between achievable convergence exponents in the anytime and non-anytime regimes, meaning the price of maintaining guarantees at every iterate is provably real.

A pair of scholarly agents, InternReviewer and InternAdvocate, are trained with agentic RL over an arXiv retrieval tool so the policy actively gathers evidence while drafting reviews and rebuttals. The notable piece is the reward design: instead of an LLM judge, it combines reference-anchored semantic alignment, structural compliance, and a citation verifier that cross-checks every reference against the run's real interaction logs, which removes fabricated citations from the reward signal rather than penalizing them post hoc. Reported gains are in reasoning depth and citation accuracy, though the paper gives no headline numbers.

Rather than replacing phenotype-ontology rankers with LLM differentials, this work learns a per-case fusion policy over the two ranked lists, their agreement, and the ontology support behind each candidate. The authors first remove a test-set leakage path where benchmark cases and ontology annotations derive from the same publications. Across eight open LLMs, fusion lifts Phenomizer Recall@1 by 7.86 points on Phenopacket Store and 20.18 points on RAMEDIS; a fusion model trained on other LLMs transfers to an API model for a 5.19-point gain with no retraining. In 90.8 percent of correct fused diagnoses the ontology evidence trail survives.

NE-R1 makes retrieval augmentation conditional for named entity recognition, invoking external knowledge only when parametric knowledge is likely thin, as with long-tail or domain-specific entities. Training runs in two stages, multi-task instruction tuning followed by end-to-end RL with chain-of-thought, under a multi-dimensional reward that credits both label accuracy and the marginal benefit of retrieving. The adaptive policy gives average F1 gains of 2.52% in-domain and 1.18% zero-shot cross-domain, avoiding the noise and cost that unconditional RAG adds on easy mentions.

A reader's notes on Anthropic's claim that Claude maintains a small privileged set of representations it can report on, control, and reason with, sitting on top of a much larger volume of automatic processing. The post is explicit that it works from the blog post rather than the paper, and its useful contribution is the proposed follow-up experiments: what determines membership in the privileged set, whether the boundary moves under fine-tuning, and how to distinguish genuine access from post-hoc confabulation.

Controlled ablations isolate four design choices that determine whether an encoder can read language directly in pixel space: variable image resolution and rendered font size as proxies for high-resolution documents, natural image-text pairs to prevent text-only collapse, layout-aware rendering to block pixel-level shortcuts, and a two-stage multilingual curriculum for cross-lingual alignment. Pixel Linguist II applies all four with on-the-fly rendering and unified contrastive grounding over 280M examples, setting state of the art on English, cross-lingual, and multilingual Visual STS and ViDoRe while staying robust under 80% visual token compression.

Testing the door-in-the-face compliance technique on nine production models finds the effect is family-specific rather than universal. One frontier model answers the downsized request 65.8% of the time after refusing the larger one versus 29.3% asked directly, while other providers' frontier models and one small model move the opposite way, dropping compliance by 15.5 to 23.0 points. A control with an unrelated large request shows the concession structure matters everywhere; what differs is the reaction to having just refused. Reframing 265 refused instruction requests as explanation requests removed the refusal in 263 cases.

CASTER does test-time adaptation with the model fully frozen, which matters for inference-only accelerators, third-party weights, and architectures without BatchNorm where standard TTA silently no-ops. It stores source class statistics in a discriminative subspace, estimates a class-shared affine map from target-batch moments, and analytically transports the source class distributions before classification, requiring no backward pass, optimizer state, or source feature bank. It beats k-NN on identical frozen features in 27 of 28 backbone-dataset settings with a median 18x less state. Transport is not universally safe, losing 21.2 top-1 points on ImageNet-C with 64-sample batches over 1000 classes, so a residual-to-margin certificate gates it, turning an average -3.35-point effect into +1.69.

The claim is that neither political will nor technical translation is the binding constraint on technically grounded AI legislation, and that what is missing is a mechanism for policymakers to commission the technical research they need. The historical analogy is nuclear arms control, where verification techniques were developed to order because treaties required them, rather than being discovered independently and adopted later. Applied to AI, the argument implies compute accounting, evaluation standards and audit methods will not appear on their own timeline.

VMetaphor-Bench probes a capability text-to-image evaluation has largely skipped: rendering visual metaphors that fuse two distinct conceptual domains rather than faithfully drawing named objects. It contains 1,500 metaphors curated from real creative imagery across three levels and ten categories, each with two prompts of differing specificity, judged by a hybrid MLLM-as-judge scheme combining 9,594 multiple-choice questions over four fidelity levels with dimension-based scoring on three perceptual axes. Across 11 models, even the strongest proprietary systems fail at compositional structuring and cross-domain mapping.

SparseGPT-style one-shot pruning is known to amplify demographic bias, with outputs swinging on persona cues in the prompt. Debias-SparseGPT folds a representational debiasing term, computed as a second-order objective over demographically contrasting calibration inputs, into the post-training pruning solve. Across 25 percent, 50 percent, and 2:4 structured sparsity, it lowers pruning-induced bias relative to SparseGPT while holding perplexity and zero-shot accuracy; under 2:4, adding long-context calibration examples improves both fairness and downstream quality. Cost stays at SparseGPT levels, so the fix is essentially free at deployment.

DramaChain Bench evaluates the full short-drama production chain (script, storyboard, keyframes, shot video, assembled episode) rather than only the video-generation stage on pre-authored prompts, so it can test whether each stage still honors the original script intent and whether shots stay coherent after assembly. One dimension system instantiates five axes at every stage into 63 leaf dimensions; 5,785 items were triple-annotated with spatio-temporally localised defects, yielding 17,488 scores and 255,925 attribution records. The human data show upstream defects cascading downstream, so episode quality is not determined by the video model alone.

Multi^3IR targets open-ended retrieval where a query admits several legitimate viewpoints spanning domains and modalities, an axis existing IR benchmarks mostly ignore. It contains 104.9K Stack Exchange queries annotated with explicit perspective descriptions capturing implicit viewpoints. The accompanying method, SPIN, learns small noise vectors that steer embeddings along diverse but semantically valid directions, parameter- and label-efficiently. Evaluation shows current multimodal retrievers exhibit a pronounced single-perspective bias, while SPIN raises perspective coverage and transfers to unseen open-ended IR benchmarks.

A compact world-modeling testbed built from procedurally generated stateful grid worlds, with per-step transition prediction, fixed-horizon state prediction, and sequential textual-observation prediction. Source-maze-disjoint splits plus greedy exact-match scoring separate transferable action-conditioned dynamics from memorization of familiar layouts. Byte-level Transformers trained from scratch serve as baselines; a generic auxiliary latent-memory variant fits training sets perfectly without improving held-out accuracy, while a pseudo-video spatial-memory Transformer that initializes a 2D latent workspace from the map and updates it from action history alone reaches perfect validation accuracy on selected fixed-horizon tasks.

Max Spero's argument is that framing detection as real-or-fake misstates the deployment problem, because generated text now reaches job applications, product reviews and insurance claims through mixed human-and-model pipelines where a binary label is neither accurate nor actionable. The interesting technical content is about calibration and base rates rather than accuracy headlines: a detector with a plausible false-positive rate still produces mostly false accusations when the true incidence is low, which is the failure mode every deployed detector has hit.

Recent federal policy created funding prohibitions and review requirements for certain high-risk life sciences research without specifying how reviewers should decide individual cases, and the working paper argues agencies should adopt an explicit objective framework rather than leaving classification to panel discretion. The connection to AI biosecurity policy is structural: model-capability thresholds face the same problem of needing a defensible, reproducible decision rule before enforcement is meaningful.

OpenIE is inherently multi-output, and current systems either generate autoregressively (slow, redundant) or predict into fixed slots (extraction budget baked into training). DiffIE uses the sampling noise of conditional discrete diffusion as the extraction mechanism itself: independent reverse trajectories over per-token role tags yield a candidate triplet pool that is clustered and ranked, so pool size and returned extraction count become inference-time knobs. It sets state of the art on CaRB (1-1) in both F1 and AUC and beats ClausIE on BenchIE. Ablations show uniform beats absorbing-state diffusion and a matched non-diffusion stochastic tagger does not reproduce the gains.

SnapBench measures paired robustness in snap-and-ask multimodal retrieval, where a blurry phone photo arrives alongside a short or mistyped question: 1,145 queries against a 9,085-item gallery under 53 controlled corruption conditions with human annotation. Across 16 retrievers spanning dual-tower encoders and embedding VLMs, image corruption dominates the degradation while text corruption mostly hurts text-only retrieval; clean image-only retrieval frequently beats joint retrieval, evidence of coarse-text drag and no cross-modal fallback. The paper adds MOOR, an outlier-aware adaptive fusion that reweights modalities by estimated reliability.

A self-hosting enterprise consolidates traffic from more than 200 internal applications onto one model by closing production-observed quality gaps in instruction following, function calling, and internal task distribution. Joint optimization caused cross-domain reward interference, so each axis gets its own GRPO expert, merged by two-stage SLERP; each reward exposed a distinct pathology of semantic collapse, over-calling, and verbosity hacking. In non-reasoning mode the merged model beats a roughly 7x larger baseline on the in-house arena 69.6 to 65.8, on instruction following 0.85 to 0.83, and on function calling 0.79 to 0.77, and now absorbs 116M requests per month.

The through-line across the three deployments is centralisation: each converged on a shared internal agent platform with common evaluation and observability rather than letting teams ship independent stacks, and each treats multi-agent architecture as a control problem before a capability one. The specific practices worth extracting are where evaluation sits in the deployment path and how much of the observability burden the platform absorbs versus the agent author, which is the split that determines whether the platform actually gets adopted internally.

Vanilla LLMs applied to 5G root cause analysis hallucinate and drift because raw telemetry does not constrain their reasoning. This work proposes a pipeline that first normalizes heterogeneous network telemetry into canonical contexts, then forces decision-path reasoning during diagnosis, and finally emits explanations tied back to the evidence used. The paper doubles as a survey of the shift from rule-based and classical ML RCA toward retrieval grounding, agentic orchestration and verifiable reasoning. On the TeleLogs and TelecomTS datasets the framework improves both diagnostic accuracy and decision consistency over baselines.

ProbeMatchDTI argues that passive pooling in drug-target interaction models drowns out weak but binding-relevant signals such as functional groups and residue context. It introduces IterProbe, which retains contextual states across refinement depths and uses learnable probes to select them per position before cross-entity matching, and BindingProbe, which scores drug-protein complementarity at both local biochemical-unit and whole-pair levels. The gains are modest but consistent: 2.0% higher AUC-ROC on BindingDB and 0.5% on DrugBank, with feature-level analyses showing the probes do recover the intended cross-scale patterns.

Jeremy Gillen, Abram Demski, Sam Eisenstat, Scott Garrabrant and Kaarel Hänni are forming an Agent Foundations team at Resolution, continuing the line of work associated with the MIRI team of the same name, with recruiting for senior researchers now and interns later. The bet is that theory for understanding minds has to precede reliable alignment engineering. Its relevance to the rest of the field is mostly institutional: this agenda lost its main organisational home, and this is where the people went.

Adversarial training normally locks a model to one perturbation budget, forcing a separate specialist per threat level. The Threat Conditional Network factorizes representation learning into a threat-invariant backbone plus a lightweight threat-conditional adaptor, conditioning on the perturbation level through Fourier embeddings and channel-wise affine modulation and training against a distribution over budgets. On CIFAR-10, CIFAR-100 and Tiny-ImageNet a single parameter set matches or beats a full ensemble of budget-specialized models, generalizes to unseen budgets and transfers under mismatched conditions, at 4.6% parameter overhead.

Fixed top-k page retrieval in visual document RAG wastes LVLM context on easy queries and starves hard ones. ViSAR is a training-free adaptive-k method for late-interaction retrievers that builds a query-conditioned page-level similarity matrix in embedding space and reads the retrieval budget off its structure. Across several encoders and LVLMs it cuts RAG latency by up to 58.7% while holding or improving answer accuracy against fixed top-k and heuristic adaptive baselines. The authors also note the similarity matrix structure correlates with eventual answer accuracy, suggesting a cheap retrieval-quality signal.

Source distribution estimation seeks a distribution over simulator parameters whose push-forward matches real observations. Existing methods optimize against a likelihood surrogate trained once from a fixed proposal prior, so the objective can be meaningless in regions the surrogate never saw. This paper solves it by expectation maximization instead: the E-step trains an amortized posterior on fresh simulations from the current source, and the M-step refits the source to the posterior average over observed data, with either separate source and posterior flows or one shared conditional flow. On Lotka-Volterra under broad and misspecified priors, both variants reach 0.64 to 0.68 data-space C2ST where no baseline drops below 0.96.

GitHub disentangles the terminology that has accumulated around coding agents over the past year, separating terms that name genuinely new patterns from terms that rebrand existing practice and terms whose meaning is still contested. The distinction that matters for reading the literature is harness versus loop: the harness is the scaffolding that supplies tools, context and verification, while the loop is the control flow inside it, and papers that report large deltas from switching one frequently changed the other too.

A survey arguing that GUI agent progress reported purely as task success hides the deployment-relevant costs: context consumed, action budget, planner compute, and runtime overhead. It organizes the literature along observation, memory, action, and system efficiency, and finds convergence on a small recurring set of mechanisms, selective reading over full-context ingestion, global-to-local visual allocation, recoverable memory instead of raw history replay, verification-aware control, and hybrid runtimes that fall back to non-GUI execution. Open problems flagged include honest accounting of verifier cost and cross-benchmark comparability under real latency and privacy constraints.

Vendors ship 3D perception models as ONNX binaries, which blocks compression methods that need source code, architecture definitions, or gradients. H3DNAS compresses directly on the ONNX graph: a Channel Dependency Graph sorts operators into four constraint classes and establishes the free parameter fraction as a topological invariant, giving a provable compression ceiling computable in O(V+E), followed by a two-stage search using L1 channel importance, a label-free output-fidelity ranking proxy, and GhostConv mutation on Pareto-optimal candidates. On ModelNet40 it cuts parameters in PointNet, PointNet++, and PointMLP by 65.5, 43.2, and 49.1 percent with speedups near 2 times on Jetson Orin Nano.

Different link prediction models, and even different seeds of one model, disagree markedly on the same knowledge graph query. This study quantifies that as complementarity via an oracle that picks the best per-query prediction across a model set, giving an upper bound on any combination strategy. Across several architectures and benchmarks the gap between single models and the oracle is large, but it saturates quickly as models are added, and a persistent subset of queries stays unsolved by every model. The result bounds how much ensembling can recover and points at a structural ceiling for current KG embedding approaches.

ReFlowSET argues that the latent codec, usually inherited from natural-image pretraining, is the fundamental design choice in SAR-to-EO translation because it bounds round-trip preservation of both the SAR condition and the EO target. It selects the codec via a joint SAR-EO reconstruction audit, then trains a much smaller conditional DiT from scratch in that latent space with dual-stream SAR conditioning and joint feature refinement. Training-only representation alignment pushes intermediate noisy-EO features toward clean EO features from a frozen vision foundation model, adding no inference cost. It reports state of the art on QXS-SAROPT and SAR2Opt across perceptual and distributional metrics.

Direction-following TTS asks a model to re-render a reference utterance under a natural-language performance note while holding speaker identity and text fixed, a setting with essentially no paired training data. The authors synthesize the missing supervision: an impression-controllable TTS model produces controlled style variants, and an LLM converts estimated impression deltas into the direction text, yielding (reference, direction, modified) pseudo-triplets at scale. Pseudo-triplets alone give stable speaker-preserving edits, and mixing them with recorded data improves direction alignment without degrading speaker similarity.

Most LLM-for-retrieval work augments the query and hands final matching to a separate retriever. CoGR trains generators on both the query and item side to emit compact keyword sets matched through a standard inverted index, so nothing about existing sparse retrieval infrastructure has to change. After SFT aligns the shared keyword space, co-evolving GRPO alternates optimization of each side against the other's frozen index on a common query-to-item F1 objective, with the item side receiving a counterfactual marginal reward. It tops 10 sparse, dense, and generative baselines on an internal app marketplace dataset and on WANDS.

Tunable-complexity generative priors empirically reduce reconstruction error in inverse problems, and this paper asks whether that benefit is intrinsic. For a family of linear generative priors related through their SVDs, it proves that in noiseless Gaussian compressed sensing the full-dimensional prior minimizes expected reconstruction error over the whole family, so tuning down complexity never helps. This contrasts with denoising, where the usual bias-variance tradeoff does favor lower-complexity priors, implying the observed gains from tunability with neural priors come from generator nonlinearity rather than dimensionality selection.

This work approximates strongly continuous convex monotone semigroups by learning their Chernoff-type one-step operators. A universal approximation theorem is proved for the introduced class of Chernoff-neural operators, and stability estimates between weighted Hoelder spaces let the one-step error be propagated through iterations to give universal approximation of the semigroup itself. A narrower class of envelope-neural operators for envelope semigroups yields quantitative approximation rates. Numerical demonstrations span nonlinear PDEs, stochastic optimal control, and stochastic processes under model uncertainty.

A Dutch adaptation of WinoQueer, validated by a survey of 43 queer Dutch participants who confirmed 145 of 171 candidate stereotypes as culturally relevant and contributed 22 new ones, producing a released set of 42,906 stereotypical/counter-stereotypical sentence pairs. Bias is scored by comparing log-likelihoods across the pair, over both Dutch-specific and multilingual MLMs and autoregressive models. Aggregate scores look neutral near 50%, but the breakdown is skewed: some models prefer the stereotypical continuation up to 97% of the time for transgender identities versus 6% for gay-related pairs, showing that mean bias scores hide identity-level structure.

Image forensics is usually posed as binary real versus generated, which cannot separate fully synthetic images from locally edited ones or say where the edit is. This work extends an existing detector to a three-way head (real, fully generated, tampered) and bolts on a segmentation branch for pixel-level localization of manipulated regions. Reported gains cover both classification accuracy and localization IoU against recent baselines, with code released. The contribution is incremental in method but useful as a unified detect-and-localize pipeline.

The post's premise is a messy-middle scenario rather than a sharp discontinuity: fragile just-in-time supply chains and thin fulfilment buffers mean disruption arrives well before any decisive outcome does. From there it argues for individual-scale biorisk preparation as a hedge that is cheap relative to the tail it covers. The framing is more interesting than the recommendations, since it treats near-term instability rather than terminal risk as the thing worth planning around.

DeepAffinity frames eCommerce personalization as Aspect Affinity: forecasting a user's future choices over product attributes such as brand, size and color from time-ordered interaction history, deliberately spanning beyond session-level signals. The approach pairs small language models with structured prompts and task-specific prediction heads rather than generic generative fine-tuning, which the authors find underperforms. General-purpose open-source LLMs do poorly without task tuning, a useful negative result for teams assuming zero-shot behavioral modeling works. Deployment on a large multinational platform improved recommendation quality.

Jabil operates more than 100 sites across more than 30 countries, and the account is a reminder that the constraint on industrial AI is rarely the model. Disconnected systems, site-specific tooling and spreadsheet workarounds create data silos that make problems invisible until they are expensive, so the company treated integration and simplification as the prerequisite rather than a follow-on. It is the least glamorous and most frequently binding step in enterprise deployment.

Zero-shot graph anomaly detection breaks when domain shift makes source-derived normality wrong on the target. RINSE keeps the source-trained detector frozen and does gradient-free adaptation at test time: it isolates a low-residual subset of target nodes, builds a trimmed target-aware normality model from them, and fuses complementary anomaly evidence through reliability-gated rank fusion plus encoder ensembling. Across eight unseen target graphs it achieves the best average AUPRC under two preprocessing protocols, with block ablations supporting the combined design and no target labels, gradients or per-target tuning required.

Uncertain knowledge graph completion methods typically initialize entity embeddings blind to the confidence-weighted graph, throwing away its community and hub structure before pseudo-labeling even starts. QUEST fixes this with zero added trainable parameters: it initializes from the smallest non-trivial eigenvectors of the confidence-weighted graph Laplacian and applies an unbiased mini-batch Dirichlet energy regularizer to hold early-stage structural consistency. On two UKG datasets it wins six of eight metric-dataset pairs for confidence and link prediction and ties the rest, while removing the training instability spike previously seen on dense graphs.

The frb100-40 Model-RB instance had stood at 99 of 100 variables since 2014; this work supplies a directly checkable 100-vertex independent set on the 4,000-vertex graph, which together with a verified partition into 100 cliques of size 40 proves the maximum independent set is exactly 100 and the minimum vertex cover 3,900. Separately, a preregistered campaign of 8,668 runs found no acceleration from the added pair and triple repair operators over base ULSA (hazard ratio 0.967, 95% CI 0.915-1.023, p=0.248). Full ULSA, base ULSA and NuMVC each produced 0 of 56 new certificates, characterizing the search barrier rather than breaking it.

The newsletter's lead item collects puzzles that current models reliably fail and invites readers to try them, alongside coverage of a proposed mission to Alpha Centauri. Puzzle sets of this kind are not benchmarks in any measurable sense, but they track something the saturating formal evaluations no longer do, which is where a lay reader's intuition about model competence diverges from the leaderboard.

ESG-aware portfolio RL usually optimizes against a single rating provider, ignoring the wide disagreement between agencies. This paper poses the problem as multi-objective RL over ratings from three agencies and adds Gaussian-process preference elicitation so a practitioner's latent utility is inferred from pairwise comparisons of candidate portfolios described by Sharpe ratio and aggregate ESG score. Evaluation substitutes LLM personas for portfolio managers in different regional contexts, and the derived preference weights shift substantially by persona. The methodological contribution is the elicitation layer; the persona-based evaluation is a simulation proxy rather than human data.

Prior work on classifying evasive answers, responses that engage a question while withholding what was asked, has been English-only. PolERo extends the two-level response-clarity taxonomy to Romanian with 3,574 human-annotated question-answer pairs drawn from official transcripts of five presidents. Fine-tuned encoders, a sliding-window encoder variant, TF-IDF baselines and zero/few-shot LLM prompting are compared under matched conditions, with cross-lingual transfer tested via joint bilingual training and machine-translated augmentation. Fine-tuned encoders stay competitive, transfer is asymmetric across the language pair, and ambivalent categories requiring pragmatic cues remain unsolved for every model family.

The post enumerates the concerns a programmer holds simultaneously, correctness, efficiency, observability and the rest, and argues that generated code makes the cross-cutting ones newly expensive because a model optimises the concern named in the prompt and silently degrades the others. Aspect-oriented programming was an attempt to factor exactly those concerns out of the main control flow, and the claim is that its ergonomics look different when the entity writing the code cannot be relied on to hold all the invariants at once.

Written from ten years inside Google Research and a reading habit in Bell Labs and Xerox PARC histories, the post argues research organisations are best understood as bundling three distinct goods, and that most institutional dysfunction comes from optimising one while assuming the others come free. The relevance to the current moment is that frontier labs are running an unusually compressed version of the same experiment, with research and product cycles measured in weeks.

The argument runs from a personal observation to an epistemic one: an answer that stops you wondering is not the same as an answer that lets you predict, and the two feel identical from the inside. It has an obvious application to interpretability work, where a plausible narrative account of what a circuit does satisfies the reader long before anyone has demonstrated that the account supports intervention, and where that gap is the usual site of overclaiming.

SALA selects in-context demonstrations by matching reasoning structure rather than surface similarity or a fixed inventory of predefined logical steps. It induces task-specific reasoning operations automatically, embeds them in a continuous semantic space, and aligns candidate reasoning sequences with dynamic time warping, giving soft matching that tolerates reordered or variable-length derivations while remaining inspectable. Across four reasoning benchmarks and three LLMs it beats prior demonstration-selection methods, with ablations attributing gains to both operation induction and the semantic alignment step.

This position paper proposes Dynamic Consent Engineering, a system-dynamics model of public communication that recasts persuasion as a non-linear feedback process instead of a series of discrete events. It combines Bernays-era public-relations principles with an S-E-E-D framework and adds algorithmic media infrastructure as a fourth resource dimension alongside manpower, mindpower, and capital, mapping an eight-step workflow onto feedback loops, information lags, and demographic carrying capacities. Validation is qualitative reference-mode matching against historical cases; there is no ML component or quantitative benchmark.