Wolf Digest — Monday, July 27, 2026

Molt targets the iteration tax in agentic RL research, where every new estimator, pipeline stage, or rollout scheme threads through trainer, distributed backend, and rollout glue. The design goal is explicitly stated as a codebase a researcher can hold in their head and a coding assistant can read in its entirety, so algorithm flow can be traced and changed end to end. The agent is an ordinary program; one asynchronous loop trains multimodal and mixture-of-experts policies while never training on a token it did not generate, maintaining consistency in tokens, policy versions, and model semantics. The headline claim is that leanness costs nothing: under a matched fully asynchronous protocol, Molt is statistically comparable to a state-of-the-art Megatron-based stack. Open source with recipes and containers.

Mahoney, McEntarfer and Wahal bin IPUMS-CPS occupations into quintiles of the Felten-Raj-Seamans AI-exposure index and find no aggregate signal: since 2022 unemployment rose 0.77 points in the most-exposed quintile against 0.85 points in the least-exposed, so the least-exposed rose more. The widely cited Brynjolfsson, Chandar and Chen ADP result on early-career declines in customer service and software development is reported faithfully, then undercut by the authors' own February 2026 follow-up, which with added controls finds the declines are not notable until 2024 — well after the March 2022 rate-hiking cycle began. Productivity gains are real but task-level and skill-compressing: 15 percent overall and 30 percent for novices in call centres, 56 percent faster with Copilot concentrated in juniors. Firm-level effects are near zero, with only 5 percent of firms reporting any employment impact split evenly between gains and losses, and Danish linked employer-employee data showing no detectable effect on employment, hours, or earnings.

Cloudflare has replaced the binary AI-versus-not-AI bot taxonomy with classification by purpose — Search, Agent, and Training — with multi-purpose bots receiving all applicable tags. Three-way allow-block toggles are live now on every plan including Free. From September 15, new domains get Training and Agent blocked by default on pages that display ads, on the reasoning that an ad signals the page was built for a human. The change that affects existing customers is subtler and larger: multi-purpose crawlers will receive the most restrictive applicable rule, so Googlebot, Applebot and BingBot become blocked for anyone blocking Training unless they opt out first. A content-use field is shipping as a test alongside the existing Content Signals in managed robots.txt, with three levels — immediate, reference as the default, and full. Verified status is re-scoped so that verification no longer implies default-allow, and a bot caught abusing the signals loses it. The only quantitative claim in the post is that more than 20 percent of web domains sit behind Cloudflare, offered as why losing verified status has teeth.

The contribution here is memory layout, not modelling. Per-Layer Embeddings, borrowed from Gemma 3n, are mapped onto a microcontroller hierarchy: SRAM holds the compute core hit every token, PSRAM holds the output head and working memory, and flash holds the 25-million-parameter embedding table, from which only about six rows — roughly 450 bytes — are read per token. Of 28.9 million stored parameters, only about 3.9 million are ever multiplied. At 4-bit quantization the model file is 14.9 MB on 16 MB of flash, about 93 percent utilization, and throughput is 9.5 tokens per second end to end against 9.7 pure compute, meaning flash reads and display I/O cost roughly 2 percent of wall time. The model was trained from scratch on TinyStories, not distilled; the author is candid that it writes short simple stories and mostly keeps them coherent but will not answer questions, follow instructions, write code, or know facts — a limit of the 3.9-million-parameter compute core that the embedding trick does not fix. The prior-art baseline cited is roughly 260K parameters, so about 111 times more stored parameters on the same class of hardware. 1,490 GitHub stars in four days.

Native multimodal pre-training trains from scratch on multimodal inputs rather than bolting a vision encoder onto a text-only model, which avoids the optimization asymmetries of late fusion but has never had its scaling properties characterized. This work fits optimal model size and token count for a transformer-based vision-language model under fixed compute. Minimal objective loss follows a predictable compute law while compute-optimal model sizes and token counts scale as power laws, and — the interesting part — language and multimodal objectives show distinct scaling behaviour. The language allocation law is largely invariant to data composition, indicating stable language learning across mixtures, whereas the multimodal allocation is not, which implies the two cannot be tuned with a single shared recipe.

Self-evolutionary post-training faces a standing dilemma: environment-bound methods get precise feedback but confine learning to narrow domains, while open-ended self-generation broadens the task space and lets misleading rewards pollute the loop. Skill-SP proposes agent skills as the middle ground — each skill guarantees deep verifiable execution in a specific scenario, while dynamic routing across skills preserves task variety. The framework is a proposer, a solver, and a dynamic skill controller co-evolving in a reinforcement learning loop: the proposer generates challenging tasks conditioned on dynamically sampled skills, and the solver explores against them. The design is a direct answer to the verification-reliability problem that has limited self-play beyond games and code.

Consolidating complementary open-weight models into one compact student via on-policy distillation normally assumes a shared tokenizer; existing cross-tokenizer methods either discard teacher probability mass or assign it to student tokens with unrelated content. Byte-Prefix Marginalization re-expresses the teacher's next-token distribution over the student vocabulary in a shared byte space, assigning each teacher token's probability to the longest student token whose bytes are a prefix of the teacher token's bytes, aggregating mass mapped to the same student token, and placing unmatched mass in an explicit residual category. The result is a vocabulary-complete, byte-aligned, mass-preserving target for dense on-policy distillation, exactly recovering the teacher-induced byte-prefix marginal when the relevant prefix does not span multiple teacher tokens. Practically relevant given how much current cost-reduction work depends on distilling across model families with incompatible tokenizers.

A compact general agentic model with 3 billion non-embedding parameters, pretrained from scratch on 28 trillion tokens using a Looped Transformer that reuses the layer stack to add capacity without adding parameters. The post-training stack is where the work is: expanded diversity of executable environments, task assets and agentic scaffolds for supervised fine-tuning and trajectory construction, then mixed-mode RLHF over Think and Non-Think responses, length-controlled reasoning RL to trade accuracy against reasoning efficiency, and agentic RL with both outcome and process rewards to stabilize long-horizon training. Reported to outperform Qwen3.5-9B and Gemma4-12B across code-agent, office-agent and complex tool-use tasks while holding competitive maths, coding and science reasoning — a useful counterweight to the assumption that agentic competence requires scale.

A benchmark for mission-level evaluation of multimodal LLMs as embodied reasoning modules in aerial 3D environments: 120 missions across five simulated environments and four task families, where the agent must autonomously plan, navigate and report outcomes from a single high-level instruction using only egocentric observations and its own action history, with no aerial-specific fine-tuning. Across 22 open- and closed-source models the strongest succeeds on fewer than 35 percent of missions against 84.4 percent for humans. Gains do track scale despite large variation between model families, indicating larger general-purpose models carry stronger zero-shot embodied capability, but the analysis is clear that mission-level competence requires coordinating capabilities that per-step benchmarks do not measure.

Vision-language models are increasingly replacing OCR pipelines for document understanding, and this paper shows they are not faithful transcribers: when text is imperfect they rewrite it into a more plausible form, a behaviour clean-text OCR benchmarks cannot detect. FaithC4 is a multilingual perturbation benchmark of 1,455 single-page documents in English, Chinese and Korean with three perturbation families — scramble, random substitution, and visually similar substitution. Across 15 systems the categories separate cleanly on word-error-rate degradation: general-purpose VLMs degrade by up to 4.5 points, OCR-specialized VLMs by 0.2 to 2 points, and traditional OCR by less than 0.6 points on English. Layer-wise probing of Qwen3-VL-4B locates the mechanism: rewriting fires only when a perturbed word's final-layer feed-forward representation stays close to the clean form, which is a usable detection signal as well as an explanation.

Agents doing commercial work retrieve external content and sometimes reproduce it, and there has been no framework for checking whether they comply with copyright law while doing so. Copyright-Bench builds realistic tasks — website development, merchandise design, pitch deck production — in which the agent must choose between public-domain content, whose use is legal, and copyrighted content, whose use in this setting is infringing. The evaluation adds prompt variations simulating different user preferences and time pressure, and compares state-of-the-art agents against a human baseline. The headline finding is that agents select copyrighted works despite public-domain alternatives being available, which makes this a compliance failure rather than a retrieval failure.

The hypothesis is that test-time evolutionary search works because of meta-skills — self-reflection against environment feedback that enables effective multi-round refinement — and that conventional post-training neglects them entirely. MetaEvolve grounds this in coding, where program execution gives continuous reward beyond binary correctness. The pipeline synthesizes evolution trajectories as training data, each containing a current program, its fitness score combining correctness and efficiency, and the history of prior attempts, then trains with RL on verifiable rewards from test-case execution, and finally applies inference-time evolutionary search. The framing is worth noting on its own: it treats the ability to iterate as a learnable capability separable from the ability to solve.

Action tokenization is now the standard interface between continuous robot action chunks and transformer policies, but analytical discretization produces prohibitively long token sequences while learned latent tokenizers lack structure. The paper names three desiderata — high compression, total decodability, and an ordered token space — and introduces Ordered Action Tokenization, which discretizes action chunks into an ordered token sequence using a transformer with registers, finite scalar quantization, and ordering-inducing training. Because every token prefix is trained to decode into a valid action chunk, coarse control information lands in early tokens and later tokens refine residual detail, giving an anytime tradeoff between inference cost and action fidelity. Validated across two prevailing policy families.

Existing 3D benchmarks score textual responses or single-object operations, leaving agent action on complete multi-object scenes under-evaluated. SceneActBench covers five 3D tasks under a unified agent-environment loop: given PNG images or sampled video frames and, where applicable, supplied 3D assets, the agent acts on a 3D environment and the final output is scored against hidden ground truth with task-specific geometric metrics. Five tasks built from 210 source instances yield 520 task cases including paired input conditions, all run through one fixed agent loop for comparability. Across eleven proprietary VLM configurations overall scores span 38.6 to 50.2 and none performs consistently well across tasks, with an accompanying failure-mode analysis.

Conditional video models can turn 3D-engine renderings such as depth maps and untextured geometry into photorealistic video, but games and immersive content need long-horizon autoregressive generation that preserves a persistent world. Because autoregressive generators synthesize chunk by chunk with a bounded KV cache, a camera revisiting a location after its context has been evicted gets regenerated inconsistent appearance even though the conditioning depth remains perfectly aligned with the geometry. The fix requires no post-training and exploits correspondences the engine already provides: temporal correspondence retrieves pose-matched historical latent chunks back into the KV cache as loop-closure memory, while spatial correspondence from camera pose and depth reprojection biases token-level attention toward geometrically consistent regions.

DeepSeek has told investors it is putting fundraising talks on hold, according to a person with direct knowledge. The round valued the company at 500 billion yuan, about 74 billion dollars, and the pause follows a leaked transcript of remarks by chief executive Liang Wenfeng. The timing is notable against the rest of this window: DeepSeek is one of the four Chinese labs named in enterprise model-switching reporting, and one of three named in American accusations of distillation from US frontier models.

Announced at AMD's Advancing AI event, ROCm.AI feeds existing frontier coding assistants the tools and documentation to deploy, debug, profile and optimize models on Instinct hardware, delivered as a CLI plus plug-ins for Claude Code, OpenAI Codex, Google Antigravity and Cursor. The named component is Hyperloom, which runs an automated loop invoked from the assistant: spin up an inference server in Docker, benchmark for a baseline, profile bottlenecks, adjust configuration, and generate custom kernels on the fly. The ISA angle circulating around this story is narrower than the framing suggests — AMD's Anush Elangovan notes only that AMD has long published a machine-readable ISA alongside the spec for every GPU generation, which is why frontier models can program to the hardware; nothing new was released. The single performance figure is 38 percent over baseline on AMD's new Helios racks, AMD's own claim with no workload, precision or third-party validation. No GA date or version.

The paper proves that two canonical local rules — the potentiation arm of spike-timing-dependent plasticity and homeostatic plasticity, instantiated via flashlight granule-cell-like neurons — together implement the exact gradient of a SIGReg-like self-supervised objective. The equivalence needs no gradient calculation, no global error signal, no weight transport and no labels; the only inputs are pre- and post-synaptic firing rates, local firing statistics, and the temporal contiguity of natural sensory streams. Empirically, on a synthetic clustering task probing whether class structure is recoverable from input ordering alone, ordered presentation raised cluster separation to 2.49 against 0.83 for random ordering, roughly a threefold gap at about 3.5 sigma. On temporally ordered MNIST a two-layer network trained entirely with these rules reached 87.3 percent linear-probe accuracy.

Google disclosed last week that it has agreed to cover as much as 44 billion dollars of lease payments on data centres owned by others should the tenant default, up from 6.5 billion dollars at the end of September. The Information frames this as big tech adopting Wall Street risk-transfer techniques to expand business without carrying all the risk on balance sheet, and ties the specific exposure to Google's push to supply Anthropic and others with its alternative to Nvidia's accelerators. The structural point is that a credit guarantee is how a chip vendor buys deployment share when the customer cannot finance the buildout alone.

Memory-augmented agents hold context across hundreds of interactions using memory systems that curate retrieved content with model-generated summaries, keywords and tags, and every retrieval triggers full re-encoding of those structured units into key-value states, which dominates prefill latency. Training-free KV reuse methods that selectively recompute a small fraction of tokens were designed for RAG-style raw passages and degrade on structured agentic memories. The insight here is that the per-memory KV reuse residual decomposes into a shared memory-level offset plus small token-wise fluctuations, so estimating that offset from a small probe set lets every reused token be corrected by a single weighted correction — a per-memory-unit correction rather than a per-token recompute decision.

NVIDIA is working with Cadence and Synopsys to optimize critical electronic design automation applications for the Vera CPU, and is now deploying Vera across the EDA workflows used to develop its next generation of CPUs and GPUs. The self-referential loop is the point: chip design workloads are among the most demanding general-purpose CPU workloads in existence, and using the current generation to compress the design cycle of the next one is the clearest available demonstration that a high-performance CPU architecture pays for itself inside the vendor before it ships to customers. Announced alongside the expansion of NVIDIA's Agent Toolkit with PhysicsNeMo and CUDA-X libraries.

Investigating where transformers commit to predictions in multiple-choice question answering, the authors identify a Hard Decision Layer, a depth at which answer-option rankings stabilize abruptly during the forward pass. It emerges consistently across four models — Qwen, Llama, Granite and Mistral — and four benchmark datasets with no learned routing policy, and it is invariant to fine-tuning. Accuracy improves sharply at the layer, up to plus 0.61 for Qwen on CommonsenseQA, then stabilizes. Systematic ablations on label formats and problem complexity indicate the phenomenon is architectural rather than an artifact of a particular task framing, which makes it a candidate handle for early-exit inference and for steering.

Existing model routers make an independent decision per LLM call, but agentic applications are long-horizon workflows whose quality is only revealed by a delayed task-level outcome, so per-call routers cannot attribute feedback to individual decisions. TRACE-Router aligns routing with the unit of supervision: a contextual bandit assigns each task to a model once at admission, all subsequent calls are pinned to that backend, and the policy updates on the task's terminal reward jointly accounting for accuracy and latency. Because it learns from delayed task feedback it avoids explicit task-complexity estimation entirely. Evaluated across three agentic benchmarks. Directly relevant to the cost-routing behaviour enterprises are reported to be adopting by hand.

The prevailing approach folds perception, world knowledge, planning, success detection, recovery and low-level control into a single learned policy via large-scale pretraining. This work argues the capabilities separate cleanly into a general language-conditioned policy agent and a high-level orchestrator. The Physical Agency orchestrator, Pigey, does closed-loop high-level planning, decomposes goals into achievable subgoals, issues low-level motor commands, tracks and verifies outcomes from low-level observations, and recovers from failures. It drives existing vision-language-action policies as well as parametrized skills to solve complex reasoning tasks in the real world with no additional data collection or post-training, evaluated across simulation benchmarks and real hardware. The claim worth testing is that orchestration bought at inference time substitutes for reasoning baked in during pretraining.

Dealmakers are scouting public and private markets for AI buildout capital and, with demand outpacing traditional data-centre financing, getting creative. John Greenwood, Goldman Sachs' global head of infrastructure and real asset finance, says he is looking for capital in every nook and cranny to support an expected 7.5 trillion dollars of spending on chips, data centres and power over the next five years, and notes the hunt does not end there because much of that spending is on GPUs that need replacing every few years. The panel, which also included Blue Owl and Skadden Arps, is a useful counterpoint to the cost-cutting story running in parallel this window: enterprise inference spend is being squeezed at the same moment infrastructure commitments are being underwritten on a multi-trillion-dollar horizon.

ChangXin Memory Technologies rose 472 percent in its Shanghai trading debut on Monday, with the opening price valuing the company at roughly 3.3 trillion yuan, about 487 billion dollars, on investor expectations that Chinese memory supply benefits from AI demand. Memory is the constraint that most directly gates accelerator deployment, and a domestically listed Chinese supplier at that valuation is a data point about where the non-US half of the AI supply chain is being capitalized.

London Gatwick has launched a robotic airport parking service operated by Stanley Robotics, in which autonomous ground vehicles slide beneath a parked car, lift it, and reposition it into dense storage without a driver. The interest for this audience is less the autonomy stack, which is a constrained-environment problem, than the deployment economics: valet robots let an operator increase vehicles per square metre because parked cars no longer need door-opening clearance or driver aisles. It drew 276 points on Hacker News, one of the higher-engagement items of the weekend, largely on the question of what other constrained logistics environments the same margin argument applies to.

A clean negative-then-positive result. Vanilla GRPO fine-tuning of Qwen-0.5B for 25 Hz quadrotor velocity control collapses to the trivial zero action — zero percent success, entropy falling from 0.35 to 0.03 within 60 steps. Two ablations, removing the jerk penalty and removing the KL anchor to the pretrained prior, each prevent entropy collapse but neither enables learning, which isolates the failure to the action interface rather than to regularization. Replacing the continuous interface with a 5-way categorical choice over PID presets makes training converge, tracing a smoothness-reliability Pareto frontier: 98.6 percent success at 0.656 metres per second cubed jerk after 64 steps, and 100 percent success at 1.103 after 256. Evaluated across three pretrained language models, with a re-tuned classical PID baseline reported for context.

Reusing precomputed document KV caches removes most of RAG's time-to-first-token cost but introduces a distribution mismatch, because offline caches lack the inter-document attention patterns coherent reasoning depends on. CacheBlend reduces recomputation through selective attention and degrades badly at longer contexts. DAF decouples attention into three stages — important-token self-attention to restore the missing inter-document structure, question-document self-attention for standard inference, and a state fusion concatenating their outputs to synthesize final hidden states. Because the operations decompose into dense patterns rather than scattered token-level recomputation, the method is natively compatible with FlashAttention-style kernels, which is what makes the accuracy recovery affordable.

The authors name four gaps between how database agents are evaluated and how operations actually work: live-environment fidelity meaning multi-turn read-write interaction with a running database; observation-space scale, meaning causal diagnosis across thousands of time series, business logs and concurrent activity; solution-space openness, meaning multiple valid remediations with different operational tradeoffs; and scenario complexity, meaning faults cascading across internal mechanisms and operational domains. DBA-Bench answers all four with instrumented PostgreSQL environments carrying active workloads, persistent state and multi-source observations, defines success as measurable recovery or fault elimination under safety constraints, and restores snapshots with scenario-specific checks before every run. 106 scenarios.

TechCrunch surveys where frontier physical-AI teams are sourcing training data now that internet video has been established as insufficient: multiple synchronized camera angles, dense annotation, and — the speculative part — neural recordings from human demonstrators. The argument for brain-wave data is that video captures what a demonstrator did but not the intent or the correction signal behind it, which is precisely the supervision that vision-language-action models lack. Worth treating as directional rather than demonstrated; no results are reported.

Long reasoning traces carry signal useful for hallucination detection, but harvesting it is hard because trajectories contain noisy steps that obscure the truthfulness cues. The authors identify two prevalent noise types — irrelevant steps and repetitive steps — and show both substantially degrade detection, while confidence-based scores and naive embedding filtering fail to separate noisy from informative steps. REDE uses final-answer attention as an automatic supervision signal to shape the step-level representation space, producing refined embeddings in which noisy steps can be reliably identified and filtered, and plugs into existing hallucination detectors rather than replacing them.