Wolf Digest — Monday, August 3, 2026

N_0-VTLA adds a tactile pathway to a vision-language-action backbone through a three-stage recipe: visuo-tactile pre-training on NeoData, a large-scale visuo-tactile robot corpus; staged integration of a predictive tactile pathway that distills contact priors into fine motion adjustments; and ALTER, an advantage-conditioned offline RL method that converts relative progress and trajectory-event comparisons into binary advantage labels for training on a fixed deployment corpus. The authors claim it is the first VLA pretrained on tactile data at scale, targeting contact-rich manipulation including deformable objects.

An audit of SULAND, the main public RGB benchmark for PFM-1 and PMA-2 surface-landmine detection from UAV and UGV imagery, found missing and false annotations, localisation errors, inconsistent visibility criteria and an inverted out-of-distribution class-ID convention. SULAND_v2 preserves the original images and splits but manually revises annotations, yielding 33,771 images and 12,433 boxes, and benchmarks 35 detector configurations across nine families. Annotation refinement alone lifts YOLOv8 in-distribution test mAP@50 by 14.6 to 19.6 percentage points, and fixing the class-ID convention raises mean OOD mAP — a reminder that in safety-critical detection, label quality dominates architecture choice.

Rather than bolting a safety filter onto a policy's final output, this modular inference framework modifies the flow-matching denoising process itself so generated trajectories are safe by construction. A smooth log-sum-exponential aggregate barrier enforces the constraint over entire action chunks with minimal compute overhead and without altering semantic intent, and the authors bound the 2-Wasserstein distance between generated and target distributions. No safety-specific dataset or retraining is required; validation covers two manipulation platforms and a 2D navigation benchmark with no measured success-rate degradation.

Critic-based RL post-training for VLA policies usually estimates value from single-frame observations or single-frame VLM latents, which mismatches the partial observability of robot control; naively feeding observation history in is exponentially expensive and still fails because scalar-return regression gives too little supervision to learn temporal dynamics. WCM diagnoses this as a state-approximation problem and builds a lightweight LeJEPA critic that jointly predicts future latent state and estimates value, so the representation is explicitly trained on temporal structure. It drops into both on-policy and off-policy pipelines.

Managed supervised fine-tuning APIs are becoming the access route to closed-weight robot foundation models: you submit data and receive a tuned policy, with no weights, gradients or training internals. That restricts improvement to pure imitation and rules out closed-loop methods. CLIFT works inside those constraints, running non-invasive closed-loop iterative fine-tuning against Gemini Robotics On-Device to produce humanoid specialists — effectively recovering some of the benefit of RL through repeated data curation rather than through gradient access.

FibVLA targets the tension between capturing temporal history and keeping VLA inference fast. It applies logarithmic hindsight sampling to both proprioceptive states and visual frames so long-term dependencies are covered with minimal redundancy, uses flow matching in the action expert to produce action distributions, and adds a Fibonacci recurrent inference schedule that generates long-range planning steps under real-time closed-loop feedback. Reported gains are in action smoothness and success rate with no retraining of the large visual encoders.

Physics simulators demand asset construction and calibration and still leave a sim-to-real gap, while video generators lack precise control over responses to fine-grained robot actions. The Boundless World Model combines initial-environment guidance, dynamic visual history and temporally aligned action conditioning for stateful autoregressive prediction of future observations, explicitly including risky and failure-prone outcomes that are expensive to collect on hardware. It is released open-source.

Cognition has signed a memorandum of understanding with the US Department of Energy to join the Genesis Mission, a national initiative the company describes as America's Manhattan Project for AI. Separately in the same window it announced a partnership deploying Devin across a cybersecurity practice serving more than 260 clients including 26 of the Fortune 500 and the top five global banks. Both were captured from the blog listing this run; neither had appeared in a prior digest window.

Text conditioning has resisted scaling analysis because diffusion loss does not scale with prompt token count. The finding here is that converged diffusion loss does scale with the amount of structured language in the prompt, quantified two ways: a white-box likelihood metric and a black-box attribute metric. Across controlled runs the converged loss decreases roughly linearly in the first and follows a power law in the second. Acting on that, the authors improve diffusability by building structured prompts with semantic and geometric annotations derived from images, and promptability by training a prompter through supervised fine-tuning, cold start and verifier-gated on-policy distillation. The system beats all evaluated open-weight models on nearly every compositional, reasoning and world-knowledge benchmark.

RLVR's reach stops where correctness stops being deterministically checkable, and the usual substitutes — human preferences, reward models, LLM judges — bring evaluation bias, judge capability ceilings and inference cost. Borrowing the pretext-task idea from self-supervised learning, RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules generate reward automatically. The instantiation, SpyRL, gives agents asymmetric information on a shared target task and has them vote to identify a designated spy; because the spy identity is predetermined, voting outcomes are fully verifiable rewards.

Safeguards decide whether to answer before seeing how the answer will be used, which is a structural problem for dual-use tasks: the same answer helps an authorised professional or an attacker, and an attacker can imitate a benign request and interaction history. Separating the capability the model releases from the evidence available about downstream use, the authors derive an exact worst-case floor on attacker assistance conditional on that evidence being copyable, yielding a trilemma between useful capability, reliable safety and open access. They then show a trusted credential — hard-to-copy information predicting actual downstream use — can complement existing safeguards, and identify the stronger condition needed to eliminate the floor entirely.

A capability assessment of humanoid and quadrupedal platforms across five axes — hardware, locomotion, autonomy, data and applications — identifying recent advances and the open challenges blocking widespread adoption. The outlook section covers ethical considerations, economic potential, policy implications and broader societal effects, and treats scientific discovery as a distinct use case rather than a byproduct of commercial deployment.

Standard diffusion and flow-matching policies couple an uninformative Gaussian prior to the action space, forcing the model to learn a complex, high-cost vector field. Temporal Policy reformulates action generation as a temporally coupled transport problem using stochastic interpolants, initialising the flow at the robot's recent history so past states are explicitly coupled to future action sequences. The data-dependent coupling reduces transport cost and produces straighter fields, which is where the inference-cost saving comes from.

ExtractBench covers 4,869 pages across 370 enterprise documents, 8 business domains and 67 document types, with ground truth built from independent-system agreement on real documents, known values for synthetic lists and human verification for forms. It is the first benchmark the authors know of to score value accuracy, record completeness at scale, grounding and measured cost together, reporting order-insensitive value F1 plus word- and page-level grounding F1. The headline pattern: commercial VLMs do well on short documents but truncate record lists on long ones, while coding agents hold accuracy at much higher cost.

Training-free vision-and-language navigation queries a multimodal LLM each step, and long horizons turn either image streams or dense maps into a memory and reasoning bottleneck. HAM-VLN makes memory decision-coupled and agent-authored: in the same model call that selects the next action, the robot records room type, objects, navigation progress and failure notes into a persistent depth-grounded world graph, keeping recent waypoints verbatim inside a bounded window while older context is compressed into the graph.

End-to-end driving models are black boxes that struggle in complex scenes, and the usual fix — bolting on a VLM for explicit reasoning — depends on pre-generated annotations that are both expensive and potentially wrong. This teacher-student framework has the teacher VLM generate logical explanations and then reflectively refine them against outcomes, so supervision is grounded in driving results rather than in human-labelled rationales, combining structured reasoning with geometric precision in the student.

RLVR broadcasts one response-level reward across every token; on-policy distillation scores each token against a stronger teacher for a dense advantage but caps the student at teacher quality. Combining them with a fixed coefficient triggers entropy collapse through two miscalibrations — a magnitude mismatch where token-level OPD advantages spike past the bounded RLVR advantage and erase it, and a temporal mismatch where sustained full-strength OPD keeps pulling the student toward the teacher. SAF applies sparsify-then-compress for magnitude and warm-up-then-anneal for time, only to the OPD advantage, evaluated with GRPO across seven math and code benchmarks on Qwen3 at 1.7B, 4B and 8B.

System prompts govern foundation-model behaviour throughout commercial products but are rarely disclosed to the public or to regulators. AISPA evaluates individual instructions along eight user-relevant dimensions, classifying each as protective or problematic, applied to 3,249 instructions drawn from 88 commercial products. Design varies enormously — some organisations average over 60 protective instructions per product while others average fewer than five — and protective coverage is wide but shallow: 98.9 percent of products contain at least one protective instruction, yet only 24 percent cover all eight dimensions. System prompts have also grown steadily longer and more protective over time.

Driving world models typically do dense prediction of future video, occupancy, bird's-eye-view representations or agent motion. Auto-JEPA argues planning only needs the scene features that affect future ego action, and predicts an intent embedding aligned with the latent representation of the future ego trajectory from visual observations, egomotion history and navigation commands. The predicted intent retrieves executable trajectories from a fixed memory, which are then ranked by a scene-conditioned scorer.

RL works well for text-to-image and single-image editing but does not transfer to multi-reference editing because no reward model captures multi-image relational constraints, and using an MLLM zero-shot forces a choice between hallucination-prone long-form reasoning and under-powered short-form judgments. EVR decomposes evaluation into distinct visual criteria; per criterion an Evaluator generates multiple candidate hypotheses and a Verifier grounds each claim in concrete visual evidence before accepting it. That yields fine-grained rewards usable for RL fine-tuning of off-the-shelf editors with no architectural change, with substantial gains over base Qwen-Image-Edit.

A teacher-student loop pairs a deterministic migrator, analyser, runner and parity gate with an agentic authoring layer that may only propose mutations, never accept them. Witness Search runs six input-space algorithms in parallel — pairwise and three-way interaction testing, Latin hypercube sampling, adaptive random testing, MAP-Elites and UCB1 — all converging within two to three branches of each other, which the authors read as a structural ceiling. When search plateaus, the authoring layer force-sets a mock-returned value symmetrically in both COBOL and Java. On a 4,114-line production-shaped program, coverage reached 135 of 142 paragraphs and 91.9 percent of branches across 166 executions with zero parity failures. The load-bearing caveat is the authors' own: the oracle is the legacy program, so the loop certifies compatibility rather than correctness, and migrated bugs are reproduced by design. No model is named.

Long-horizon scientific analysis has lacked process-supervised environments over real scientific data, so agents get rewarded on final claims rather than on the analytical steps that produce them. SciDisco compiles hypotheses, datasets, hidden evidence graphs and verifiers into task environments where progress is checkable mid-interaction, synthesises verifier-filtered multi-turn demonstrations grounded in a DAG, and uses DiscoPO to assign turn-level credit to actions producing verifiable evidence. SciDisco-14B is reported at state of the art on hypothesis-driven data-analysis benchmarks.

Apollo Global Management analysts scored 321 US occupations against Anthropic's Economic Index, which measures the share of an occupation's tasks observed being performed with Anthropic tools, and compared Bureau of Labor Statistics wage data for 2022 against 2024. The most AI-exposed occupations show an average 6.7 percent decline in real wage growth after 2023, with no detectable effect on employment levels. The distribution is skewed: service workers average a 24.3 percent decline in earnings growth, the bottom quartile of earners 10.7 percent, and there is no significant effect among the highest paid. The exposure measure is one vendor's product telemetry, the window is short, and the data contains visible confounds — radio DJs show a 52 percent real wage collapse at low exposure while administrative law judges gained 17.5 percent at moderate exposure.

Rubric-based RL is known to suffer from Unexplored Criteria that no rollout satisfies and therefore receive no gradient; the usual fix injects rubric text as rollout guidance, creating a train-inference mismatch. This paper names a second failure mode: Suppressed Criteria, satisfied by some rollouts but whose signal is lost because scalar reward aggregation assigns them non-positive aggregate advantage. Over 57 percent of samples show this throughout training, averaging 1.8 suppressed criteria each. Criterion-Distilled Policy Optimization addresses both without the guidance mismatch.

Rule-based multi-robot coordination depends on predefined task models and bespoke decision programs, which do not survive complex semantic instructions or heterogeneous platforms. D-VLC pushes instruction interpretation, task decomposition and role assignment into a decentralised vision-language layer so each robot reasons about objects, regions and spatial relationships locally, without a central planner holding a global model of the environment.

RGB observations carry no explicit geometry, so visuomotor policies learned from them break under camera perturbation. RayViT represents camera geometry as a Plucker ray map, patchifies it into ray features, and uses gated cross-attention to build a ray-conditioned class token; ray features are added as dense positional embeddings and the ray class token replaces the original ViT class token. The base weights stay pretrained, so the cost is a small conditioning module rather than a new encoder.

Translation with Thought trains a resource-rational policy that modulates inference between intuitive and deliberate reasoning according to domain difficulty: supervised fine-tuning on difficulty-aware long chain-of-thought traces distilled from DeepSeek-R1 and rewritten for reasoning economy, then RL with a hybrid reward balancing translation quality against reasoning cost. Across 15 benchmarks spanning in-domain and out-of-domain settings, 3 seen and 59 unseen languages, TwT at 7B and 14B outperform much larger state-of-the-art reasoning models while cutting token usage by 32 to 60 percent.

A walkthrough of NVIDIA-affiliated work on human-to-humanoid transfer arguing that pure motion imitation fails because the human demonstration does not encode the contact forces and balance corrections the robot actually needs. The video links to the project page for the underlying method; the episode is a summary rather than a primary release.

Analysing 22,500 agent trajectories across GAIA, SWE-bench and Multi-Challenge, the authors define an Evaluative Dissonance Index quantifying how much LLM-as-a-judge systems mistake structural proceduralism for semantic truth when a candidate mimics consensus. A logistic meta-evaluator isolates the syntactic triggers of evaluator capture at 0.8779 ROC-AUC, and zero-shot leave-one-domain-out transfer at 0.7482 mean ROC-AUC indicates the vulnerability is domain-agnostic rather than an artifact of any one benchmark.

Two Google product retractions landed within a day of each other and both reached the Hacker News front page: the AI Studio app was cancelled after roughly 800,000 preorders, and a separate Earth AI generator was withdrawn about a day after launch. Announcements came through the respective product accounts rather than a formal blog post, so the stated reasoning is thin; the pairing is what drove the community reaction.

Researchers built a life-cycle model as a normative benchmark, had 1,000 adults write their own advice-seeking prompts, and simulated ages 22 to 89 following GPT-5.2, GPT-5.6 and Gemini 3 Flash recommendations. Advice was broadly sound — saving during working years, drawdown in retirement, diversified equity, de-risking after 45 — but handled shocks poorly, leaned on rules of thumb and let portfolios drift instead of rebalancing. Prompts written by men, by more financially literate users, or by users with prior AI experience produced roughly 5 percent more wealth near retirement, compounding to about $50,000 less at age 60 for women and less literate users; about two-thirds of the gender gap traces to how prompts are written and one-third to the model changing advice when the same prompt is labelled as coming from a woman. Models also volunteered specific providers users never mentioned — Vanguard in 6 percent of responses, iShares in 3.4, against under 0.4 percent of prompts naming either.

Mainstream mesh generation serialises a mesh into tokens and decodes autoregressively, which is slow and accumulates error. Meshy T2 uses a vertex-set VAE that encodes a mesh into one continuous latent token per vertex and decodes vertices, edge connectivity and face winding order in a single pass — no vertex quantisation, no welding, artist-authored topology preserved. Generation is a coarse-to-fine cascade: an image-conditioned voxel flow sketches an occupancy scaffold, then a mesh flow populates it with per-vertex latents conditioned on image, scaffold and a requested vertex budget, giving explicit face-count control.

The second instalment of the China travel series covers how AI safety work is organised and practised inside Chinese labs and institutions, following a first part on the broader research and deployment landscape. Useful primarily as first-hand reporting on institutional structure and evaluation practice, an area where secondary coverage tends to rely on translated policy documents rather than direct conversation.

Long chain-of-thought reasoning brings redundancy accumulation, context overflow and error anchoring, and the authors argue the real bottleneck under a bounded window is neither trajectory compression nor test-time control but the absence of a reusable intermediate interface that can replace discarded history. They also identify a specific failure of outcome-reward long-chain RL: when the model has not solved the task and the window is nearly exhausted, the final-answer reward rewards premature guessing. ThinkReset instantiates interface writeback and reset in text space and optimises post-reset continuation success directly.

Most agent memory systems spend extra LLM calls generating intermediate records and mediating retrieval, adding recurring token and latency cost while merged or omitted details obscure the original evidence. Zero-Mem asks whether structured memory access needs generation at all: it preserves raw interaction traces and organises them two ways — an entity-context graph for cross-interaction connections and a temporal hierarchy for conversational locality — weighing both views per query. Deterministic calibration discards conflicting evidence, and only the final reader invokes a model.

Evaluating autonomous research systems is unresolved because the output is a paper, not a score. This benchmarking protocol runs an automated peer-review system built on frontier models, assessing originality, scientific rigour, clarity and significance, and applies it to Sakana AI v1 and v2, CycleResearcher and Data-to-Paper. Each framework ran on a consistent set of 15 research proposals published by a commercial autonomous-science company, producing 60 papers for comparison under identical review conditions.

I-MLFFs replace explicit stacks of neural network layers with self-consistent fixed-point equations, which lets intermediate representations be reused across successive molecular-dynamics timesteps to warm-start force evaluation. The result combines the compute footprint of a shallow single-layer force field with the representational capacity of a deep network, and the efficiency gain is architecture-agnostic — demonstrated across invariant, equivariant Cartesian and other graph neural network classes — because it comes from coupling force prediction to trajectory integration rather than treating them separately.

A federal judge declined xAI's request to enjoin Minnesota's statute banning applications that let users generate nudified images of real people, allowing the ban to take effect while the company's lawsuit proceeds. The ruling is notable because it leaves a state-level restriction on a specific generative-image capability operating against a frontier lab, rather than against a downstream app developer.

Where LoRA adds a low-rank update to weight matrices, LARA reads the hidden state at a small set of layers and adds a low-rank correction back into the residual stream, leaving base weights untouched entirely. It matches LoRA at equal parameter count on a code fine-tuning task and on DPO. The practical difference is that because adaptation is a frozen base plus a residual, LARA exposes an inference-time scale that interpolates smoothly between base and adapted behaviour — graded control that weight-space adaptation does not offer, and a route to composing several adapters.

Scalar reward models are efficient and probabilistically interpretable but lean on superficial cues that fail out of distribution; generative reward models reason their way to robustness but produce natural-language scores lacking numerical flexibility. Recent hybrids combine both through off-policy multi-task learning, which does not guarantee the generated reasoning actually helps the scalar head. This work trains the latent reasoning trace end-to-end against the scalar prediction so the trace is optimised for the downstream reward rather than alongside it.

Multi-model systems increasingly share contexts, retrieved evidence and dialogue histories, but key-value caches are model-specific, so every model re-prefills or separately stores the same context. MoT translates a source model's context cache into a target model's cache space using multiple translator modules to capture diverse source-target mappings, rather than relying on a single projection path or a global shared latent space — the scaling argument being that one projection cannot cover architecturally distant pairs.

Agents acting on partial observations face belief-state inference, objective misalignment and planning under uncertainty, and conditioning on full or summarised action-observation histories drags redundant context into every decision. NeSyFS represents the belief state as a knowledge graph and feeds triplets to every module: a fast-thinking module for reactive action, a slow-thinking module doing uncertainty-aware planning structured on twisted sequential Monte Carlo, and a reflection module to correct objective drift.

Reddit shares dropped 23 percent following results attributing a slowdown in user growth to AI assistants answering questions that previously drove search traffic to the platform. The item matters less as a market move than as a datapoint on the referral-traffic mechanism now visible in a public company's reported numbers, since Reddit content is simultaneously a major licensed training and retrieval source.

On-policy distillation's rollout cost is dominated by a few long responses delaying batch completion, and existing accelerations cap length with fixed budgets or absolute teacher-student agreement thresholds that do not track learning progress across models or training stages. Adaptive FastOPD expands the horizon only when four teacher-student signals indicate learning near the current boundary has plateaued and the current horizon is sufficiently utilised, making the schedule progress-aware rather than pre-set.

Existing agent benchmarks measure static code generation, paper replication or final-answer correctness, none of which test whether an agent can interpret results and act on them. AgentHPOBench provides 30 executable machine-learning tasks across seven research categories, each starting from a validated baseline run, after which the agent performs several sequential interventions while observing accumulated configurations, metrics and logs — turning hyperparameter optimisation into a sequential-decision evaluation rather than a one-shot generation task.

Small language models are attractive for agentic deployment on latency, cost and on-device privacy grounds, but they cannot absorb noisy tool-use supervision the way large models can, making data quality the binding constraint. Data Turnstile takes user-defined API specifications and decomposes multi-turn tool-use interactions into constrained stepwise generation with validation and error-feedback loops, giving explicit control over API diversity, conversation complexity and output correctness. Released open-source.

Federated fine-tuning of foundation models runs into resource asymmetry: the institutions holding the most valuable domain data cannot host billion-parameter models. Parameter-efficient tuning, pruning and distillation each trade away one of full-model memory reduction, architectural self-containedness, or representational fidelity. FedSLM decomposes with SVD to produce self-contained client models whose low-rank subspaces form nested manifolds, so clients at different capacities occupy compatible subspaces of one server model.

Solutions occupying larger volumes in parameter space, quantified by Boltzmann entropy, are known to generalise better than those conventional optimisation reaches — the high entropy advantage. This paper asks whether the advantage extends to robustness, meaning retention of learned knowledge when the model is subsequently trained on new information. Using grokking in modular arithmetic as a controlled setting with a noise-injection protocol, it separates apparently-grokked solutions from true equilibrium ones and finds the latter substantially more resistant to forgetting.

NeuroSynth borrows complementary learning systems from biology — rapid hippocampal encoding alongside slower cortical consolidation — and implements it as a dual-pathway continual RL architecture separating rapid task acquisition in a plan pathway from long-term retention in a habit pathway, combined with replay and knowledge distillation. The target is the standard failure where training on new tasks overwrites earlier competence.

Financial deployments combine retrieval, proprietary data, tool use, orchestration logic, monitoring and human escalation, yet evaluation stays model-centric — benchmark scores and one-off qualitative reviews treated as evidence of readiness. Drawing on industry experience validating generative AI applications inside financial institutions, the authors argue for validation evidence spanning data, model design, retrieval and generation performance, agent behaviour, governance and implementation as a precondition for production approval.

Latent world models depend on the quality of the learned latent distribution, and LeWorldModel regularises toward an isotropic Gaussian with the Epps-Pulley objective — whose corrective gradients, the authors show, vanish rapidly for isolated tail samples, leaving heavy-tailed deviations uncontrolled. QQWorld swaps in a quantile-quantile matching objective that aligns projected latent samples with rank-matched Gaussian quantiles, keeping gradients alive in the tails, plus a cross-batch variant enlarging the ranking pool with detached prior-batch samples. Across four control environments it improves average planning success and yields thinner latent tails.

Quantization's memory savings are well characterised; its effect on latency and throughput under realistic, controlled orchestration-level workloads rarely is. This study covers two translation model families across five models from 1.7B to 22B on a single A100 or H100, and finds that pairing a document-chunking strategy with W4A8 or W8A8 quantization improves the latency-throughput Pareto curve across a wide operating range rather than at a single batch size.

RadixArk and Google Cloud announced a partnership to bring SGLang to TPUs so developers can run the same serving stack across hardware backends. Captured from the LMSYS blog listing this run alongside two adjacent posts — Blackwell-native MXFP8 and per-token NVFP4 reinforcement-learning recipes in Miles, and a cleaner quantization stack in SGLang — all published just before this window opened.

A vendor-run audit of 531 stored scans across 458 domains reports that only 10 of 193 sites were ever named across 5,978 buying-intent assistant answers, and that just 1.9 percent of answers named the business being asked about — Gemini 2.9 percent, ChatGPT 1.7, Claude 1.6, Perplexity 1.6. Separately, 8.9 percent of sites block at least one AI crawler, and the blocking is aimed at training rather than retrieval: 38 sites block GPTBot while only 4 block OAI-SearchBot. Caveats are heavy — the corpus is self-selected from a vendor selling AI-visibility services, the headline rests on 193 sites, and the structured-data figures are an acknowledged under-count because the detector missed JSON-LD graph containers until 2 August.

Ultra-high-definition restoration has to aggregate spatially recurring degradation cues while preserving local structure, and the usual cost controls — downsampling, window partitioning, cluster-based token reduction — attenuate edges and textures because they discard what shared aggregation represents poorly. CoDe-SSM decouples the two: a global cluster pathway models aggregated context while a separate pathway carries clustering residuals, so fine structure is retained rather than reconstructed.

Automated pose estimation and individual re-identification are letting field researchers run cognitive experiments on wild primate populations at sample sizes that manual video coding could not support, shifting the constraint from analyst hours to camera coverage. The methodological point that generalises is that the bottleneck in observational animal cognition has been annotation throughput rather than experimental design.

Most agent benchmarks use bounded tasks with immediate success criteria, which cannot detect whether an agent preserves purposeful behaviour across a long horizon. MerchantBench builds a persistent seller-side e-commerce environment where actions constrain future choices, feedback arrives at heterogeneous delays and incoherent behaviour accumulates measurable cost, across product sourcing, listing and pricing control, and cash-flow management.

Unifying text-, image- and video-conditioned generation in one model needs a VLM's multimodal understanding wired into a pretrained video diffusion transformer, and existing approaches either inject features from a few manually chosen VLM layers or jointly train architecture-matched understanding and generation streams, which blocks reuse of heterogeneous pretrained backbones. MoRoute learns the routing instead, dynamically selecting which hierarchical VLM representations feed the diffusion transformer per condition.

TAPR reformulates a user prompt into a task-optimised one with the explicit objective of improving downstream performance, trained with GRPO where the reward comes from LLM-as-judge evaluations of both the rewritten prompt and the resulting task output. Reported gains span question answering, summarisation and arithmetic reasoning, with the practical framing that non-expert users should not have to carry prompt engineering themselves.

Detectors for AI-generated images degrade sharply under post-processing that shifts low-level statistics. RAID exploits the asymmetry between generated and camera-captured images under bit reversal of the input representation, using the difference in detector response before and after the transform as the discriminative signal rather than relying on a single forward pass. Timely given the EU labelling obligations that took effect this window, where watermark absence has to be backstopped by post-hoc detection.

TechCrunch's Equity podcast discusses Sam Altman publicly calling on the industry to pace the rate of AI development, and what that framing means coming from the chief executive of the lab with the largest commercial deployment. The item is commentary on a public statement rather than a policy announcement, and the segment does not report any concrete commitment attached to it.

Polymer informatics is usually split across separate models for forward property prediction and inverse structure recommendation, which makes end-to-end design loops brittle. UniPolymer puts prediction, recommendation and evaluation in one framework so candidate structures are scored by the same representation that generated them, closing the loop without a handoff between independently trained components.

Coverage of solo operators reaching seven-figure revenue by delegating functions that previously required staff — support, content, bookkeeping, basic engineering — to AI tooling. The article reached the Hacker News front page largely on the strength of the pattern rather than the sample size; the underlying reporting is anecdotal rather than statistical, and it sits alongside the Apollo wage findings elsewhere in today's digest as the optimistic reading of the same automation trend.

A working novelist sets out, step by step, where generative tools would plausibly slot into a professional fiction workflow and why each insertion point is rejected — covering drafting, continuity tracking, research and copy-editing separately rather than treating the question as a single yes or no. It pairs directly with today's market-dilution paper on AI-written self-published fiction.

GQ-FSL combines federated split learning with quantized activations and gradients, optimising explicitly for energy consumption on edge clients rather than treating energy as a downstream consequence of communication rounds. The framing matters as federated deployments move onto battery-constrained hardware where the binding limit is joules per round, not bandwidth.

InMyStyle is a privacy-first single-user system that adapts small models to rewrite AI-edited text toward an individual's writing style without any instruction prompt at inference. Local helper models build paired training examples from the user's own documents, then LoRA adapters are fine-tuned on bases from 0.5B to 7B parameters, with length-aware generation budgets and automatic chunking. On 219 evaluation pairs from a scientific-paper corpus the composite score plateaus at 0.69 across every model size under both greedy and sampled decoding — the negative result being that capacity is not the bottleneck for this task.

Model attribution for generated video normally requires training a classifier per generator, which does not survive new model releases. This retrieval-driven approach matches a query video against a reference bank of known-generator samples in a representation space, so adding a generator means adding references rather than retraining — the practical property that matters when the generator population turns over every few months.

The twenty-third open-artifacts roundup argues against the consolidation thesis that many observers treated as inevitable as training costs rose by orders of magnitude each year. Laguna S2.1, Inkling and Kimi K3 are used as the counter-evidence: capable open-weight releases continuing to arrive from a widening rather than narrowing set of labs, with enough production utility that the gap to closed frontier models is now argued in terms of specific workloads rather than in general.