Wolf Digest — Monday, July 13, 2026

Ireland's Central Statistics Office reports datacentre electricity consumption rose 10% in 2025 to 7,663 GWh, from 6,973 GWh in 2024, while every other class of customer grew 2%. That puts datacentres at 23% of national metered consumption, up from 20% in 2023, 14% in 2021 and 5% in 2015. They now consume more than urban dwellings (18%) and more than twice the rural household share (9%). The 10% rise happened while an effective moratorium on new grid connections around Dublin was in force for nearly all of 2025; the Commission for Regulation of Utilities lifted it in December. New connections above 10 MW must now provide generators or battery systems capable of matching their draw and feed power back to the grid on demand. Ireland has more than 80 datacentres for a population just over 5 million. Notably, the CSO figures do not attribute the growth to AI specifically.

Ploy's migration writeup is the most instructive vendor post of the week because the headline result — 2.2x faster, 27% cheaper, $3.06 to $2.22 per build, 8m00s to 3m42s, visual score 0.936 to 0.970 — is the number they got after fixing their harness. The raw first run had the new model looking about 50% more expensive. Two findings generalise. First, tool-argument pathology: on a schema with 25 top-level parameters and one required, GPT-5.6 sent all 25 on 6,635 of 6,635 calls, inventing plausible values, while Opus 4.8 did so 4 times in 2,898 calls. An invented <code>offset: 0</code> meant 52–64% of its file reads silently returned empty with <code>success: true</code>. Neither prompt directives nor OpenAI's strict mode fixed it; rewriting optional properties as required-but-nullable at the provider boundary took empty reads to 0% and cut tool calls ~30%. Second, caching: GPT-5.6 dropped partial-prefix matching, so their 29K static prefix cached at 0% across new conversations, and every uncached prompt also pays a 1.25x cache-write surcharge. Workspace-scoped cache keys took first-call hits from ~0% to 83.7%. About a third of their initial cross-model failures traced to harness assumptions rather than model behaviour. Small n (11 and 10 builds), single workload.

The argument is that large-scale text-to-video generation is to vision what next-token prediction was to language: the pretraining objective that supplies the spatiotemporal priors, vision-language alignment and scalability a generalist visual model needs. GenCeption turns a pretrained video diffusion backbone into a feed-forward perception model steered by text instructions, and reports state-of-the-art across depth, surface normal and camera pose estimation, expression-referring segmentation and 3D keypoint prediction — matching or beating specialist models including DepthAnything3, SAM3, VGGT-Omega and Sapiens. The ablation that carries the claim is that the video-generative backbone outperforms alternative pretraining paradigms on the same downstream suite, which is the part that would need independent replication before the analogy to next-token prediction is earned.

James Evans (Chicago) and collaborators at the Beijing National Research Center analysed 41.3 million English-language papers from 1980 to 2025 across biology, chemistry, physics, medicine, materials science and geology — deliberately excluding computer science and mathematics to avoid circularity — and used an NLP classifier to isolate roughly 311,000 AI-augmented papers. Scientists who adopt AI publish about 3x as many papers, receive nearly 5x as many citations, and become team leaders one to two years earlier. Collectively, the same research occupies a smaller footprint in a high-dimensional knowledge space, clusters more tightly around data-rich problems, and generates weaker follow-on engagement networks. The pattern holds across early ML, deep learning and the generative wave, and Evans says it is intensifying. His framing: the tools reward speed and scale but not surprise, and the danger is homogeneity rather than slowdown. Bowen Zhou (Shanghai AI Laboratory) offers the counterargument that the narrowing is an artifact of fragmented, siloed AI-for-science tooling and that integration would expand discovery; Evans replies that the binding constraint is incentives, not architecture. The study is observational, and IEEE Spectrum reports no causal-identification strategy, confidence intervals or a quantitative measure of the footprint shrinkage. Published in Nature on 14 January and resurfaced on Hacker News this weekend.

A hybrid Mamba-Transformer MoE activating 3B of 30B parameters per token, pretrained on roughly 27 trillion tokens with German deliberately up-weighted, and built end-to-end on the German Industrial AI Cloud — a sovereign HPC-scale cluster operated by Deutsche Telekom in Munich. The hybrid design keeps the inference cache near-constant as context grows, which is the throughput argument for long-context, high-concurrency serving. It matches dense 14–27B models on aggregate English and German benchmarks, posts the best code aggregates in both languages among 17 open base models, and beats every European sovereign baseline compared, including ones with far larger active parameter counts; among fully open models it leads Olmo 3 32B and Apertus 70B. Release is under permissive open-access terms including weights, selected intermediate checkpoints and full per-source data accounting — which, in the week the open-weights policy debate sharpened, is the part that matters.

An adversarial trainer can plant backdoors in a large class of deep feedforward networks such that the backdoored and honestly trained models are close in total variation distance even given full white-box access to every weight. The backdoor grants its holder invariance-based adversarial examples for every input — mapping distant inputs to unusually close outputs — while, without the key, generating any such example is provably impossible in polynomial time under standard cryptographic assumptions. The result establishes a fundamental power asymmetry between whoever trains a model and whoever uses it, and it is the strongest formal argument yet that weight-inspection is not a sufficient supply-chain defence. Theoretical with preliminary empirics.

The linear representation hypothesis underpins probing and activation steering, but prior work asks whether linear concept directions should exist after training, not how they emerge during it. This paper derives exact solutions for the full trajectory of what it calls abstraction in a minimal linear network, yielding three analytic principles: data and target geometry jointly fix the end-of-learning abstraction; abstraction improves with depth; and initialisation scale controls the maximum abstraction reached during training. Extending to nonlinear networks, it analyses how the choice of nonlinearity shapes the dynamics. The practical hook is that if initialisation scale bounds achievable abstraction, the interpretability of a model is partly a training-time decision rather than a property to be discovered post hoc.

Builds a transformer from bounded, named units that read as fuzzy set operations rather than dense activations, and shows that legibility has to be actively pressed for during training — with a failure mode. A crispness penalty intended to sharpen a bounded operator into a decisive detector instead collapses it into a dead constant; the identity E[v(1-v)] = mu(1-mu) - var explains why (the penalty is a variance minimiser blind to the difference between a live detector and a constant) and names the fix, a per-channel variance floor written as a loss. A learned per-unit fraction then retires the hand-set reserved-GELU partition of prior work: given the choice, the model keeps no unit as pure GELU and routes 87% of its load-bearing computation through crisp operators. 78% of feed-forward operands and 50% of attention value channels end up as crisp-and-contextual detectors, with per-head legibility rising from 18% in shallow layers to 78% in deep ones.

An OpenAI-compatible endpoint that pools GPUs and memory across machines on a peer-to-peer mesh: a request either runs locally, routes to a peer that already has the model loaded, or — in the mode codenamed Skippy — splits a model too large for any single box across several machines by layer ranges, streaming activations from stage to stage. This is pipeline parallelism, not tensor parallelism, which is the right choice over consumer networks and also the source of its limits. Every node boots an iroh endpoint whose public key is its identity and only network surface; iroh handles hole-punching, NAT traversal and relay fallback to open direct authenticated QUIC connections, with three ALPNs including a dedicated latency-sensitive one for inter-stage activation transport. The catalog ships 40+ models from half-a-billion parameters up to 235B MoE, and the binary is about 18 MB. Worth being clear about what the post is not: it contains no measured throughput, latency or bandwidth figures, and no comparison to Petals, exo or vLLM, and it does not discuss stage stalls, straggler nodes, activation bandwidth cost, or how peer compute is verified.

A systematic study of reward hacking when multimodal LLMs are aligned with text-only or weakly grounded rewards, across safety VQA, chart VQA and stress tests, varying reward design, data ambiguity, model scale from 2B to 32B, and algorithm (GRPO, RLOO, DAPO). Outcome-only rewards reach a 48.1% reward-hacking rate. The paper's sharpest instrument is Newly Rewarded Failure Rate, which counts failures among samples whose proxy reward improved over the SFT baseline; NRFR exceeding the hacking rate shows RL is manufacturing new failure modes, not just inheriting them. Scale attenuates but does not remove the effect — the 32B model still degrades 54.9% under outcome-only rewards. GRPO is consistently the most resistant; RLOO stays vulnerable; DAPO improves substantially from 2B to 8B. Answer-aware and visual-evidence rewards improve the oracle trend at every scale.

Rich Sutton's short 2024 essay resurfaced on Hacker News this weekend, and it reads differently in 2026 than it did on publication. The one-step trap is the assumption that an agent's predictions can all be one-step, with long-horizon predictions generated by iterating them — a one-step transition model rolled out like physics or a simulator. Sutton grants the grain of truth (perfectly accurate one-step predictions do compose perfectly) and then makes two independent objections. Statistically, if the one-step model is imperfect, errors compound and all bets are off. Computationally, in a stochastic world the future is a tree of possibilities, each of which must be imagined and probability-weighted, making the complexity exponential in prediction length and therefore infeasible regardless of accuracy. His prescription is temporally abstract models built from options and general value functions. The line with contemporary bite: one-step models are 'hopeless, yet extremely appealing, and are widely used in POMDPs, Bayesian analyses, control theory, and in compression theories of AI.' He never names a model or a lab.

Current pipelines convert visually rich sources — documents, web pages, typeset equations, page layouts — into plain text before pretraining, discarding the visual structure. This paper runs a systematic study of unsupervised visual pretraining directly on visual documents without text extraction, and finds that across multiple backbones and benchmarks, visual pretraining on the same underlying corpora consistently outperforms text-only pretraining for language intelligence. The controlled-corpus design is what makes it interesting: it isolates the representation from the data, so the gain is attributable to preserving layout and figure structure rather than to seeing more content.

Uses adversarial suffix attacks as a probe of representational alignment rather than as an attack per se. Activation-Guided GCG replaces output-based objectives with losses targeting the model's internal refusal direction, and the informative negative result is that suppressing refusal globally across all layers and positions works better than targeting any single layer-position pair — evidence that safety representations are distributed across the forward pass rather than causally localised to one site, which complicates the tidy low-dimensional refusal-direction picture. Soft-GCG, a Gumbel-Softmax continuous relaxation of the discrete suffix search, achieves a 33x speedup over standard GCG while improving attack success. Across scales, smaller models remain vulnerable while larger ones resist more.

A replication study of emergent misalignment — the finding that fine-tuning on narrow misaligned data abruptly produces broadly misaligned behaviour, reversible by limited realignment. Running controlled cycles of misalignment and realignment while tracking behaviour and LoRA representations, the authors reproduce the phenomenon but find both directions are highly sensitive to superficial dataset characteristics: the apparent rapid realignment largely disappears once response-length differences are controlled for. They also find the previously reported mechanistic signature — representational phase transitions in LoRA space — does not consistently correlate with behavioural misalignment across training. The conclusion is not that emergent misalignment is false but that the current evidence is weaker than claimed and the evaluation protocols need to control for surface artifacts.

Not a typeface but a video: letters are drawn in dots identical to the background, legible to humans only through motion, so a paused frame or a screenshot reveals nothing. Every generation also embeds a decoy message, on the reasoning that an agent with local code execution analysing dot motion will surface the decoy first. Passed to two frontier models, both read the decoy and failed to recover the real message until told the exact technique to look for; an earlier reasoning model spent 19 minutes on a still frame and hallucinated a message that does not exist. The framing is the interesting part: the 2013 anti-OCR typeface ZXX, called surveillance-proof at the time, is read correctly by a current model on its fastest tier in a single prompt. The author is candid that this is not security — a code-executing agent can decode the motion, real hiding requires encryption, and the defence exploits a specific and temporary bottleneck, namely that multimodal models are image-based and sample video as individual frames. He expects a video-native model to read it directly. He also notes it is fairly hard for humans to read. Evidence is screenshots of individual chat sessions, not a controlled evaluation.

Terminal benchmarks mostly score final outcomes on tasks that finish in minutes, which yields sparse reward and hides partial progress. This benchmark contributes 46 long-horizon tasks across nine categories — experiment reproduction, software engineering, multimodal analysis, interactive games, scientific computing — each decomposed into finely graded subtasks with a reference solution or simulation engine, enabling dense intermediate reward and partial credit. Tasks typically require hundreds of episodes and minutes to hours of execution, stressing long-horizon planning, long-context management and iterative recovery. The dense-grading design is the part worth borrowing: it makes 'how far did the agent get' a measurable quantity rather than a binary.

Within-class variance in language-model representations is usually read as incomplete neural collapse. This paper argues it is allocated information storage obeying a law, and takes the unusual step of retracting the authors' own earlier simplex equiangular-tight-frame claims via a one-line centering identity. Across 14 models spanning a 100x parameter range, macro-category structure carries only 4–12% of representational variance while within-token context carries 79–91%. Theoretically, token-level weight decay penalises a category in proportion to its type count rather than its occurrence mass, reducing next-token prediction to an imbalanced K-class problem whose optimum orders category norms by type count; a converse floor, proved for binary categories, forces within-category dispersion to be at least proportional to the conditional mutual information I(token; context | category). Identity dispersion tracks that information across every model and partition tested — even across models, where one model's information predicts another's dispersion.

Fine-tuning an LLM on new facts reliably produces memorisation without downstream reasoning use — an accuracy gap plus a temporal lag the authors formalise as the knowing-using gap. Using a self-patching intervention that identifies activation locations where relocating a representation rescues a failed generalisation case, they support a knowledge-circuit misalignment hypothesis: the memorised representation exists internally but is not routed to computation-effective layers. The diagnostic pays off — a simple heuristic derived from it recovers 58–75% of the oracle headroom on generalisation failures, cross-domain.

A polemic against the AI 2040 scenario document, and beneath the rhetoric there is a specific technical thesis worth extracting: intelligence is a bottleneck for some things but not the binding constraint on changing the physical world. Hotz's argument from hardware experience is that shipping a device of cell-phone complexity is dominated by supply-chain latency, spec deviations, parts failing after twenty minutes, and chips warping in the reflow oven — none of which more or better tokens remove. His latency-floor examples are concrete: three weeks on a boat versus air freight; three months to fabricate a chip, with humans barely in the loop. He grants the long-run substrate argument, that machines or hybrids are plausibly the successor species and that space suits them better than us, while maintaining there is no hard takeoff because there is no physical trick that intelligence unlocks. He also makes an agent-alignment case for local models with a clean framing: an assistant hosted by a company with hotel partnerships is not the assistant you want booking your hotel, and 'if you can't kick it, it's not aligned with you.' The post contains no benchmarks, no data, and a large amount of deliberately provocative material.

Existing studies treat coding-agent failure as a final outcome; this one treats it as a temporal process with an onset, an evolution and a recovery window. The authors collected 3,843 execution trajectories from seven frontier models across three scaffolds (OpenHands, MiniSWE, Terminus2) on Terminal-Bench, filtered to 1,794 complete valid trajectories, and manually annotated over 63,000 execution steps, deriving 14 findings on failure occurrence, root cause, recovery and cross-system consistency. The headline: failures are predominantly driven by epistemic causes — the agent's model of the environment diverging from the environment — rather than by capability limits, which relocates the fix from the model to the harness's feedback loop.

Block-sparse attention is the hardware-friendly answer to the KV-cache read bottleneck, yet leading open-weight models still use dense or fine-grained selection. Taking DeepSeek's Native Sparse Attention as the representative method, the authors reduce block selection to ranking blocks by attention mass — the sum of a block's attention scores — and prove that retrieving the highest-mass blocks lets block-sparse attention match dense quality. Computing exact mass requires reading every key, so selection becomes the problem of approximating mass from a compact summary. A cumulant expansion shows why existing selectors falter: they are confined to a first-order approximation of that mass. COBS extends to higher cumulant orders.

An endorsement experiment in which four models evaluate the same international economic and security policies after each is randomly described as supported by the United States, the European Union, China, or Russia. In the numeric-only condition, GPT-5, Claude Sonnet and Gemini rate China- and Russia-endorsed policies substantially lower than identical policies endorsed by the US or EU; DeepSeek is the main exception. Requiring a short written justification alongside the score leaves the Western/non-Western gap intact for GPT-5 and Claude Sonnet, attenuates Gemini's penalties, and sharply activates the China and Russia penalties in DeepSeek. The justifications show Western endorsement treated as a credibility cue while Chinese and Russian endorsement is treated as a cue for data security, sovereignty or surveillance concerns. Directly relevant to anyone using models as policy-document evaluators.

Second-order post-training quantization methods, GPTQ included, build their objective purely from input activation statistics, implicitly assuming every output channel contributes equally to layer-wise reconstruction. KronQ brings gradient covariance in: under a Kronecker-factored Hessian approximation the quantization loss depends jointly on activation and gradient covariances, which the method exploits twice — bidirectional incoherence processing that extends the usual input-side random rotation to the output dimension, and a new inter-layer mixed-precision sensitivity metric driven by gradient and activation Hessian traces. The headline case is 2-bit weight-only quantization on LLaMA-3-70B, where the assumption GPTQ makes is most costly.

The news event is thin — a single product-manager job posting in San Francisco for family, caregiver and older-adult experiences, with no product announced and no comment from OpenAI — but the Sensor Tower data TechCrunch obtained is not. ChatGPT's share of users aged 35 and over rose to 31% globally in Q2, from 26% a year earlier, while the 18-to-24 share fell from 34% to 29%. In the US, nearly one in four smartphone-using parents used ChatGPT during the quarter, up from 16%. Across assistants, the 25-to-34 cohort is 40% of app audiences for Claude, Gemini and ChatGPT alike, and 33% for Copilot; Copilot skews oldest, with 20% of users over 45 against 11% for ChatGPT — but ChatGPT is adding older users fastest, up three points year over year while Claude and Gemini declined. Reach among US smartphone-using parents: Gemini 32%, ChatGPT 24%, Claude 4%, Copilot 2%. Separately, a Family Online Safety Institute survey of over 4,000 families found 27% of US parents said their child used generative AI in the past week while 38% of children said they had. The regulatory backdrop is stated plainly: OpenAI faces multiple lawsuits from parents alleging ChatGPT contributed to harm to their children, including cases involving suicide, and has shipped parental controls, sensitive-conversation routing to reasoning models, and an optional Trusted Contact alert feature. All demographic figures are Sensor Tower estimates of mobile app audiences, with no methodology or margin of error disclosed.

Vision-language-action models inherit semantics from pretrained VLMs, but heavy robot-data post-training and architectural surgery reshape the backbone so much that it becomes hard to say what the VLM actually contributed. CLAP takes the minimal-change path and identifies the core obstacle as output-distribution mismatch: predicting actions as bare numeric token sequences drags generation away from the pretrained language distribution and erodes the capabilities you were trying to keep. The fix is to prepend each numeric action sequence with a natural-language action description, causally conditioning action-token prediction on a language plan without touching the architecture. With single-epoch fine-tuning, 2B CLAP reaches 90.8% on LIBERO, up 14.9 points on VLA-0, and improves robustness on LIBERO-PRO under language, object and spatial perturbations.

Differentiable logic gate networks learn gate types but conventionally fix the wiring. This work makes the connections trainable too, maintaining a probability distribution over candidate connections per gate input pin and selecting the highest-merit one while gate types are learned in parallel. The efficiency result is the point: 98.92% on MNIST with two layers of 8,000 gates, and 98.45% with a single layer of 8,000 — roughly 50x fewer gates than fixed-connection LGNs for comparable accuracy. Stability up to ten layers required a high learning rate, straight-through estimators and trimming constant-output gate types. Directly relevant to FPGA and ASIC deployment where gate count is the currency.

Efficiency work on vision-language models has concentrated on cutting visual tokens, implicitly assuming visual processing dominates energy cost. The first systematic energy profile of on-device VLM inference — five models, three architecture families, four input resolutions, two platforms (RTX 3070 and Jetson Orin NX) — overturns that. Average inference power turns out to be a model-intrinsic constant, varying under 5% across resolution, image complexity and prompt type, which means all energy variation across inputs comes from inference time, not power draw. And each output token costs 11 to 39x more wall-clock time than each input token, because of the compute-bound versus memory-bound asymmetry between prefill and decode. Output token count is therefore the dominant driver of both latency and energy on the edge, and visual-token pruning is optimising the wrong term.

Test-time training treats the test context as a training example for instance-specific adaptation, but applying it to a full long context is prohibitively expensive and adapting on randomly sampled spans injects noise — most spans in a long context are irrelevant to the question. The authors' preliminary study makes the sensitivity concrete: on LongBench-v2, TTT on random spans hurts performance while TTT on oracle spans substantially improves it. S-TTT closes the loop by having the model identify the evidence spans it should learn from before adapting on them, which is the obvious move and evidently an effective one.

On edge devices, MoE inference is bottlenecked not by FLOPs but by weight swapping between slow storage and fast memory, because consecutive tokens keep activating different experts. Existing fixes are system-level caching heuristics or post-hoc router fine-tuning; both leave the root cause untouched during pretraining. StickyMoE adds a differentiable routing-consistency loss that penalises abrupt expert switches between adjacent tokens, encouraging the router to hold an assignment across semantically coherent spans. No architectural change, one hyperparameter, and — unlike post-hoc methods — expert representations and routing decisions co-adapt from the first training step. Small-scale results: up to 60% reduction in expert switch rate for under 4% perplexity degradation, Pareto-dominating post-hoc fine-tuning on the quality-locality frontier.

On-policy distillation is notoriously unstable and high-variance. TOP-D dynamically constructs a proximal teacher, and the authors establish that this inherently controls gradient variance, providing a global convergence analysis alongside a monotonic improvement bound. Empirically it improves training stability, sample efficiency and final performance on mathematical reasoning, and it adds zero computational overhead relative to standard OPD — which, given how much of the current frontier post-training stack leans on distillation from a stronger teacher, is the kind of drop-in that gets adopted quickly if it replicates.

Diffusion and flow models for protein design almost all require the sequence length to be fixed before sampling, which is awkward because optimal length is usually unknown and tightly coupled to designability. Generalized Poisson Flow learns the rate function of an inhomogeneous generalized Poisson process by minimising its negative log-likelihood, giving variable-length generation with population-level recovery guarantees for the joint multimodal distribution and a KL upper bound between data and generated distributions. Evaluated across structure and sequence design, motif scaffolding and peptide co-design, spanning Euclidean, categorical and Riemannian modalities; it improves structural designability in unconditional design and achieves the best distributional fitness for sequence design.

Generated video supplies useful motion priors for manipulation, but visual plausibility is not physical executability: a generated clip lacks metric geometry, grasp grounding, kinematic feasibility and execution-time feedback, so replaying its trajectory is unreliable. GenVid2Robot samples task-relevant semantic anchors from the real first RGB-D frame, tracks them through candidate generated videos, and verifies whether the resulting 2D motion can be explained by the first-frame anchors under a sparse relative SE(3) model — accepting only geometrically consistent motion and discarding the rest. The framing is the contribution: generated videos as uncertain hypotheses subject to a verification gate, rather than as demonstrations.

Exploits the rotation-equivariance of omnidirectional representations — where rotation is an implicit geometric transformation — to simplify camera trajectories into translations under fixed headings, addressing long-range memory in panoramic world models via Dense Panoramic Ray-Conditioning and Geometry-aware Memory Augmentation, trained in three stages. The accompanying World360 dataset combines real panoramic UAV footage with simulated AirSim360 clips specifically to test physical consistency under large spatial variation and diverse illumination, which existing panoramic datasets are too static to probe. Models, code and dataset promised.

Generated code frequently fails after deployment for structural rather than logical reasons: an endpoint references a configuration key never declared, an import targets a package in no registry, a new route omits the authentication guard every sibling endpoint has. Each patch is locally valid and globally incoherent, and standard CI rarely surfaces it. The paper formalises structural coherence as consistency invariants over graph representations of repository artifacts — import, call, dependency, configuration, schema, resource, control-flow and routing graphs — and gives an eight-category failure taxonomy that separates defects specific to LLM generation from those merely amplified by it, plus a hybrid verification framework that delegates to mature static analysis where it already works. The auth-guard example is the one that should worry people.

A position paper with a sharp diagnosis: reasoning, coding, theorem-proving, tool-use and long-horizon research all share a structural limitation — the representational frame is fixed and supplied in advance. The conceptual vocabulary, the space of admissible solutions, and the success criteria are all given, so the system searches within a space rather than altering it. The authors name two gaps. The vocabulary gap is the difficulty of inventing and stabilising new representational primitives rather than recombining existing ones. The verifier gap is the difficulty of judging a new primitive's value when its payoff only becomes visible after subsequent work builds on it. Pairs naturally with the Nature metascience result on narrowing discovery.

Starts from a measurement failure: trained human raters cannot agree on whether a personalised assistant is acquiring traits that harm its user, with a Fleiss kappa of 0.074. Three small models sharing a hyperbolic substrate are offered as an instrument. A 146M behavioural auditor trained from scratch detects the compliance gap those raters cannot, at 90.7% binary-compliance accuracy, and a linear read-out of its frozen representation detects companion-induced sycophancy, dependence-fostering and confabulated memories on generator families unseen in training — AUROC 0.804 under style-controlled leave-one-generator-out evaluation, against 0.721 for a frontier zero-shot judge on the same items. A small specialised auditor beating a frontier judge on its own held-out distribution is the claim to check.

A Planner-Worker-Verifier architecture in the mould of recent agentic proving systems: a Planner maintains a compact Whiteboard scratchpad plus an unbounded Repository of intermediate findings and decomposes work across parallel Workers, with Lean 4 providing automatic formal verification of every generated proof. The system's value is less the ProofNet numbers than the reproducibility story — automatic formal verification means ablations are quantitative rather than judged — and an interactive terminal mode that lets a human monitor and steer the proof search, borrowing the human-AI synergy that already works in interactive code generation. Fully open source.

Agentic workloads inside operating-system workflows push inference onto laptop-class SoCs, where cloud offload carries reliability and privacy costs. Recent laptop silicon ships NPUs tuned for energy efficiency, but mapping attention onto them is awkward because of architectural diversity and explicit data-movement programming models. STEEL gives a dataflow formulation of prefill attention that exploits spatial parallelism and on-chip memory, and — the part that is actually novel — attacks the load imbalance the causal mask induces via sparsity-aware pipeline placement across the NPU array, cutting synchronisation overhead. Evaluated on the AMD Ryzen AI 9 HX 370.

Decomposes each ARC-AGI-2 task into iterative perception, hypothesis generation, symbolic execution and reflective refinement under strict test-time and hardware constraints. A perceptual grounding agent builds object-centric scene graphs from raw grids, a latent program policy proposes diverse DSL programs, a symbolic executor verifies candidates against the demonstrations, and a reflective agent turns failures into feedback for the next turn — all communicating through a shared differentiable blackboard scheduled by a learned meta-controller. The combination of structured program search with adaptive multi-turn correction is the standard recipe for this benchmark; the blackboard and meta-controller are what distinguish it.

A 183.2M-parameter non-autoregressive conditional flow-matching Diffusion Transformer operating in the frozen continuous latent space of AudioVAE2 (16 kHz encode, 48 kHz decode), so capacity goes to text-to-latent mapping while inheriting high-quality 48 kHz reconstruction. Notable for what it removes: no phonemizer, no grapheme-to-phoneme frontend, no discrete speech tokenizer — end-to-end from a 92-symbol Turkish character vocabulary. Parallel denoising predicts the whole latent sequence at once over a predicted duration. A two-stage post-training recipe of single-speaker voice locking plus short-utterance coverage targets the two production failure modes. Band-matched WER 8.0%, CER 3.0% on the accompanying benchmark.