Wolf Digest — Saturday, July 11, 2026

The paper names a failure mode it calls behavioral state decay: in long-horizon tasks, decision-relevant state — task requirements, environment facts, prior attempts, diagnoses, open subgoals — gets buried in or pushed beyond the context window and stops influencing decisions. The intervention is a separate memory agent that runs alongside an unmodified action agent, maintaining a structured memory bank from the recent trajectory and deciding, at each step, whether to inject a memory-grounded reminder or stay silent. Treating memory as an active intervention rather than passive retrieval is the substantive move. The module is plug-and-play with existing harnesses and improves pass@1 on both Terminal-Bench 2.0 and tau-squared-Bench for weaker and stronger action agents alike, with gains reported above eight points.

Vidu S1 is a real-time interactive video generation model in which a user steers a digital character with voice instructions at any point during generation. Built on TurboDiffusion and TurboServe, it produces 540p video at up to 42 FPS on ordinary consumer GPUs and supports unbounded-length generation without the blurring, drift, or structural distortion that usually accumulates in autoregressive video rollouts. Users can supply their own reference images — real people, anime, pets — and select voice timbres. The interesting engineering claim is the combination of streaming interactivity, long-horizon stability, and consumer-hardware throughput in one system; a playable demo is live.

GitHub swapped Copilot code review's bespoke exploration tools for the shared Unix-style grep, glob, and view tools that power the Copilot CLI harness, expecting a clean upgrade. Benchmarks regressed: reviews cost more and caught fewer issues. The old tools were not thin wrappers — they returned matched lines plus surrounding context automatically, which suited earlier models that made few tool calls and pulled in context poorly. The fix was not to revert the tools but to rewrite the instructions around them for how a reviewer actually reads a diff, which flipped the regression into roughly 20% lower average review cost at unchanged review quality. The diagnostic signal was that the agent made a similar number of tool calls but spent more of them on relevant evidence instead of repeatedly widening the search. Notably, the same instruction style did not help in the CLI, where exploration is part of the job — shared tools scale only when instructions and benchmarks match the product.

A comparative study that expresses softmax attention and four recent recurrent linear-attention architectures — DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2 — in a common recurrent-memory notation, making explicit how they differ in expressivity, memory decay, erase and write control, training throughput, and implementation complexity. Experiments center on 350M-parameter models trained for 15B tokens, with optimizer and learning-rate sweeps, hybrid-versus-pure stack comparisons, sequence-length runtime measurements, larger DeltaNet runs at 1.3B and 3B, and downstream evaluations. The value is the shared formalism and the apples-to-apples training budget: most linear-attention claims in the literature are made against inconsistent baselines, and this is the rare paper that fixes the budget and reports where the trade-offs actually land.

Importance sampling is what makes modern LLM RL sample-efficient, and clipping is what keeps it from diverging — but the paper argues clipping does structural damage. Formalizing a notion of Probability Capacity, the authors show that conservative clipping prematurely truncates the update budget for correct-but-low-confidence reasoning paths, which is precisely the exploration you want to preserve. Their proposed objective, Unbounded Positive Asymmetric Optimization, restructures the update by anchoring the negative side while leaving the positive side unbounded, and is presented as a plug-and-play drop-in for existing RL frameworks rather than a new algorithm family.

Speculative decoding buys its speedup from idle compute during memory-bandwidth-bound, small-batch inference; at large batch, inference becomes compute-bound and fixed-K speculation actively regresses below vanilla. Cohere's dynamic speculative decoding chooses K per step from the model-hardware interaction — raising it when bandwidth-bound, lowering it when compute-bound — so speculation degrades gracefully instead of hurting. Reported results: roughly 23% faster than fixed-K speculative decoding at batch size 128 and 256, 7.5% faster than vanilla at 128 and 1.8% at 256, where fixed-K regresses. The engineering, not the idea, was the hard part: making dynamic K compatible with vLLM's async scheduler, which schedules step T+1 on CPU while step T runs on GPU using only counts and placeholders, and with full CUDA Graph, which now captures multiple valid K values per batch size so runtime K changes still hit a captured graph. Both changes are upstreamed to vLLM. On the Command A+ MoE the gains flatten, which the authors attribute to an EAGLE head trained for a single timestep being reused across K > 1.

Cursor is developing a general-purpose AI agent aimed squarely at Anthropic's Claude Cowork, according to The Information. Two people familiar with the project say work began after Cursor started leasing compute from SpaceX's AI unit in April. The strategic read is straightforward: the coding-assistant category is converging on general agentic work, and the companies that own the developer's editor surface are unwilling to cede the broader knowledge-work agent to the model labs. It also confirms that the compute relationship between Cursor and SpaceXAI runs deeper than a supply arrangement.

Zero-shot context extension is the dominant deployment path for open-weight checkpoints, because agentic traces and repository-level coding routinely push inputs an order of magnitude past the pretraining window. Existing methods fix a single RoPE rescaling factor up front, forcing a choice between short-context fidelity and long-context stability. Jet-Long pairs a local RoPE-faithful window with a long-range window whose rescaling factor adapts to the current sequence length, recovering the base model exactly at short inputs while extrapolating cleanly at long ones. An inclusion-exclusion attention merge and an on-the-fly RoPE correction keep the two windows consistent. No fine-tuning required.

DARPA has selected more than 120 organizations to compete in its Lift Challenge next month, with millions of dollars in prize money on the line. The target is a vertical-lift unmanned system that carries four times its own weight — today a specialized helicopter manages roughly one-to-one, and commercial drones are considerably worse. Program manager Phillip Smith frames the objective as cost per pound per mile: scaling current designs up makes cost explode, which is why battlefield drones remain small and payload-limited. A four-to-one ratio would let a much smaller, cheaper airframe carry the same heavy payload, which is what makes contested-logistics resupply economically feasible. The effort sits under the Department's Drone Dominance initiative.

Inference-time scaling for text-to-image has drifted toward guided search methods that verify and steer trajectories at intermediate denoising steps, and those methods are compared by number of function evaluations — a metric that counts denoising passes but ignores verifier overhead. The authors show that once you evaluate on wall-clock time instead, simple Best-of-N already matches or beats several guided-search techniques, which suggests compute is better spent on broader exploration than on repeated intermediate verification. Flash-BoN takes that seriously: it generates a large pool of cheap draft candidates by stacking three complementary acceleration knobs, then selects. A useful correction to a benchmarking convention that has been quietly flattering the more complicated methods.

Chain-of-frame reasoning is the video analogue of chain-of-thought: the reasoning unfolds across temporally connected generated frames rather than through text tokens. The problem is that video generators are trained on general video corpora with no supervision for this behavior. OpenCoF supplies OpenCoF-17K, a reasoning-video dataset spanning eleven task families, and Wan-CoF, a fine-tune of Wan2.2-I2V-A14B, to test whether diverse temporal supervision induces the behavior. Across four video reasoning benchmarks Wan-CoF posts considerable gains over the baseline, which is the first reasonably controlled evidence that chain-of-frame is trainable rather than incidental.

In a Friday memo, Elon Musk told Tesla staff to move to Grok, the model from SpaceXAI, wherever possible, citing Grok 4.5's lower token costs relative to competitors. It is a small item with an outsized signal: internal enterprise standardization on a single model is now being decided on price rather than capability, which is exactly the dynamic the week's releases set up. It also further entangles Tesla, SpaceX, and xAI operationally at a moment when Grok 4.5 is posting genuinely competitive results on long-horizon coding benchmarks.

The Defense Department's new 25-page Post-Quantum Cryptography strategy, published in June, commits to updating the Cybersecurity Maturity Model Certification with quantum-resistant algorithm requirements, and extends PQC to access control, zero trust, and software development platforms. The broad deadlines are that every DOD system must support PQC by the end of 2030 and employ it before the end of 2031. Enforcement timing for contractors is unspecified, which is the practical problem: experts quoted say the defense industrial base is likely unprepared, and there is no clear signal yet on when compliance begins to bite. Harvest-now-decrypt-later is the underlying threat model.

Scientific ideas inherit mechanisms, repair known limitations, and recombine earlier work — and no current benchmark tests whether AI systems can follow that inheritance structure. IG-Bench represents each paper or proposal as a set of minimal, typed, evidence-grounded Idea Genome objects, with a GenomeDiff aligning them to record inheritance, mutation, loss, external import, and novel insertion under six evolutionary dynamics. It ships 1,961 golden lineage traces, 1,085 curated Idea Genome objects, and 920 pairwise GenomeDiff records across ten scientific domains, supporting both an exam-style evaluation and lineage-grounded idea generation. Notable as an attempt to make 'novelty' an operationalized, checkable property rather than an LLM-judge vibe.

Agent benchmarks mostly rely on sandboxed environments and single-turn evaluation, and their scenario-based task taxonomies mix several model capabilities inside one category, which makes failures hard to attribute. UniClawBench reorganizes evaluation around five foundational capabilities — skill usage, exploration, long-context reasoning, multimodal understanding, and cross-platform coordination — and runs proactive agents in dynamic real-world settings rather than a sandbox. The capability-driven decomposition is the point: it turns 'the agent failed' into a statement about which competence was missing.

Meta has removed the Muse Image feature that let users generate images by @-mentioning public Instagram accounts as visual references, without notifying the referenced user. It shipped earlier this week alongside Muse Image's launch and drew immediate backlash; the company's Friday blog post says the feature 'missed the mark' and is no longer available. Reporting notes the reversal came amid scrutiny from users and talent agencies including CAA. The predictable-abuse failure mode here is the same one that platforms have repeatedly failed to anticipate with image models, and shipping it with no notification to the person whose likeness is being referenced was the specific error.

A new CSET report by Igor Mikolic-Torreira and Emelia Probasco examines where AI-enabled decision support actually fits into how military organizations plan, decide, and operate — deliberately looking past the targeting use case that dominates the discourse. The central finding is that data accessibility, not model capability, is the binding constraint on adoption, and the report closes with recommendations for scaling AI capabilities across the force. Worth reading alongside the Pentagon's ongoing consolidation of autonomy programs: the institutional plumbing, not the models, is where the friction lives.

The benchmark landscape splits awkwardly between symbolic causal-reasoning tests with no realistic data analysis and data-analysis benchmarks with no principled causal generating structure. CausalDS closes the gap: each instance is a scene built from a sampled structural causal model, with generated observational data and a synthetic natural-language story grounded in a realistic domain. Because the SCMs are systematically generated rather than drawn from a curated pool with templated variations, contamination and memorization are much harder. It tests the thing that actually matters for a data-science agent — whether it can recover causal structure from data it is handed, not whether it can recite Pearl.

Most generative drug design conditions on a target protein or on general molecular properties and ignores that the same target behaves differently in different disease contexts. DrugGen-2 conditions on both disease ontology and target protein sequence, fine-tuning a pretrained GPT-2 on approved drugs linked to their diseases and targets, with supervised fine-tuning followed by GRPO against reward functions for chemical validity, novelty, diversity, and predicted binding affinity. Evaluated on five targets relevant to diabetic nephropathy. The disease-conditioning is the contribution; the base model is deliberately small.

Tom Uren's Seriously Risky Business column traces a chain from a Supreme Court decision — holding that the President acted lawfully in removing an FTC commissioner without cause, and that the for-cause removal statute was unconstitutional — to the EU-U.S. Data Privacy Framework. Max Schrems argues the ruling means no U.S. executive authority is genuinely independent, which would undercut the European Charter requirement that data protections be subject to control by an independent authority; the European Commission's implementing decision names the FTC specifically. More than 3,600 U.S. businesses operate under the current framework, and Section 702 collection from Europe depends on those data flows. Not everyone buys the argument — a Dechert privacy partner calls it 'a stretch' — but Schrems has twice invalidated predecessor schemes. The column also reports, separately, that CISA is now using Anthropic's Mythos model to audit government software for vulnerabilities, and that the audits have already surfaced many.

Energy Department CIO Dawn Zimmer told FedScoop the agency is expanding model options in its Joulix suite but will not add for the sake of adding: 'There hasn't been a huge demand for Perplexity,' and 'we're not probably going to add Grok at this point.' The detail matters because federal availability and federal usage have decoupled — Perplexity was the first AI platform to sign a direct GSA Multiple Award Schedule deal, holds FedRAMP Low, and is in pilots at Justice and Labor, yet demand inside DOE has not materialized. Grok has a FedRAMP High push backed by Agriculture. DOE's broader model expansion grew out of the Defense Department's dispute with Anthropic.

Extending single-target video object segmentation to multiple targets is usually done by replicating the single-target pipeline per object, which drops frame rate linearly and gives unbounded latency as the target count grows. SAM-MT restructures SAM2 so that explicit per-target queries run in parallel alongside a shared global-context representation, using decoupled masked attention to keep identities from bleeding into each other and a sparse memory for stable temporal evolution. The result is real-time multi-target segmentation with latency that does not scale with the number of objects.

Arabic LLMs overproduce Modern Standard Arabic because dialectal data is scarce, which sets up a clean interpretability question: where are dialect features encoded, and can they be manipulated at inference time? The paper runs two complementary probes. A neuron-level analysis identifies sparse neuron populations encoding dialect-specific features and shows that amplifying or suppressing them steers generation toward a target dialect. Because dialect features turn out to be entangled at the single-neuron level, the authors also apply distributed vector-space steering directions. The interesting finding is the tension: dialect is sparse enough to find, but distributed enough that neuron-level intervention alone is not the cleanest control surface.

'aria' is a native, dependency-free runtime that executes the full Stable Audio 3 text-to-music pipeline on ordinary GPUs, CPU-only machines, and a Raspberry Pi 5 — no Python, no deep-learning framework underneath. The core contribution is a quantization study: running at lower numerical precision to fit tight memory budgets, saving memory in place rather than adding to it. Because the runtime owns every internal tensor, it also exposes activation steering cheaply. Quality cost is measured on three independent axes — prompt adherence, audio quality, and taste preservation — each compared against the model's ordinary sampling variation, which is the right control and one most quantization papers skip.

On TechCrunch's Equity podcast, Clem Delangue argues the pattern is now consistent: companies begin on frontier APIs and migrate to open models as they scale, because inference cost dominates once volume is real. Hugging Face is used by roughly half the Fortune 500. He frames the open-versus-closed question in the wake of Anthropic's halted Fable release and says his concern is concentration — a handful of companies ending up in control of the whole stack. The argument lands differently this week than it would have last week, given how much near-frontier capability just became available at a fraction of frontier pricing.

DIU released the public criteria for future Defense Innovation OnRamp Hub locations, following a solicitation open on SAM.gov from July 7 to 31. The authority comes from section 913 of the FY2026 NDAA. Selection criteria include ecosystem strength measured by the EDA Innovation Intelligence Index, proximity to private-sector technology ecosystems relevant to modernization priorities, workforce availability, evidence of community commitment, existing military presence, and proximity to test and evaluation assets. Notably, hubs must source 10% of funding from local, state, and private sources, with a stated preference for full self-sustainment within five years. A stakeholder webinar is set for July 20.

Diffusion transformers are too heavy for mobile image-to-video, and the demand is specifically for cinematic camera effects — bullet time, dolly zoom, slow motion. CineMobile stacks three compressions: distillation-guided pruning to a compact model that retains the capabilities needed for those effects, diffusion distillation plus reinforcement learning to collapse sampling into a 4-step generator, and hybrid post-training quantization. It is a straightforward efficiency paper, but the target — a narrow, well-specified capability rather than general video generation — is what makes the aggressive compression viable.

Germany's armed forces equipment office signed a contract in June with MBDA Deutschland and Rheinmetall Waffe Munition to develop a full high-energy laser weapon system — reconnaissance through target tracking and engagement — with operational fielding expected in 2029 and a contract value in the mid three-digit millions of euros. The demonstrator has already logged 28,000 nautical miles aboard the frigate Sachsen from the Baltic to the Mediterranean and fired more than a thousand shots at air, sea, and land targets over a year of testing. The stated driver is counter-drone defense, where a per-shot cost near zero is the entire argument. Separately, the French-German Research Institute of Saint-Louis conducted its first outdoor railgun test in June.

The Download's lead is a mainstream write-up of Anthropic's Jacobian lens work — the tool the interpretability team used to isolate what it calls J-space, a hidden region containing words related to the response the model is working on but may never produce. That paper was covered here on Wednesday; the newsletter's contribution is reach, not new technical content. The genuinely new items in the roundup are two: Tencent is leading a deal to unwind Meta's two-billion-dollar Manus acquisition and is in talks to become the Chinese startup's largest shareholder, after Beijing ordered Meta to unwind it; and humanoid robots have performed teleoperated gallbladder removals on live pigs, described as a world first. The newsletter also notes Meta has begun charging developers for a paid tier of Muse Spark.

Recovering video from sparse event-camera streams is a regression-versus-generation trade-off: regression blurs texture, generative models drift over long horizons. LongE2V fine-tunes a pretrained video diffusion model to handle reconstruction, prediction, and frame interpolation jointly, which buys data efficiency and perceptual quality. The stability machinery is where the work is: autoregressive unrolling with adaptive context switching to fight temporal drift over very long sequences, reencoding alignment with cross residual correction for bidirectional consistency during interpolation, and event voxel density augmentation for robustness across sensor resolutions.

Offline motion generators give precise text and kinematic control but cannot run at interactive rates; online methods run in real time but lose controllability and struggle with complex text semantics and long-horizon goals because of limited context. ARDY is a streaming framework using a hybrid representation — explicit root features paired with a latent body embedding — to keep precise trajectory control while remaining cheap enough to generate. Relevant beyond animation: this is the same control-versus-latency trade-off that shows up in humanoid whole-body controllers.

Canvas360 is a two-stage framework — geometry-aware pretraining, then task-specific fine-tuning — for panoramic style transfer, inpainting, outpainting, and editing. The dataset is the substantive contribution: Canvas360Dataset supplies a million high-quality paired panoramic samples, which is what the in-context panoramic setting has been missing. On the modeling side, parallel depth generation, velocity circular padding to handle the seam, and a similarity loss regularizer push the model toward geometry-consistent representations that capture equirectangular distortion rather than fighting it.

Most MRI super-resolution treats the problem as a deterministic map from one fixed low-resolution input to a high-resolution target, which ignores the acquisition physics: spatial resolution and signal-to-noise ratio are coupled, so any given low-resolution scan is one realization among many possible acquisition trade-offs. PhyMRI-SR reformulates the task as identifying the optimal resolution-SNR configuration first and super-resolving from there, which makes resolution a dynamic rather than fixed quantity. To handle resolution-heterogeneous inputs the authors adapt 2D Gaussian Splatting to MRI reconstruction.

The Labor Department is soliciting public comment ahead of launching a nationwide survey measuring how Americans actually use AI in their work and time. The reason to care is measurement: nearly every claim about AI's labor-market effect currently rests on vendor telemetry, self-selected surveys, or job-postings data, none of which support the population-level inference people keep drawing from them. A federal time-use instrument would be the first widely trusted denominator.

In zero-shot compositional action recognition, models are supposed to recognize novel verb-object pairs from previously seen primitives. The paper diagnoses a specific cheat: models predict the verb from the labeled object class rather than from temporal evidence, and sparse compositional supervision plus verb-object learning asymmetry actively encourage it. Diagnostic metrics show existing methods overfitting to training co-occurrence and underusing temporal verb cues. Their fix, RCORE, adds co-occurrence prior regularization with explicit supervision for unseen compositions. Another instance of the week's recurring theme: benchmarks that look solved because the shortcut was available.

Junjie Yan, founder and CEO of Chinese AI developer MiniMax, told employees in an internal memo he will take no salary until the company achieves AGI, and will allocate some of his personal shares toward employee incentives. The equity component is the part with actual mechanism behind it; the salary pledge is signaling. Worth logging mainly as a datum on how Chinese labs are competing for talent against much better-capitalized rivals.

Peaked circuits are quantum circuits engineered so that one output bit string dominates the distribution, which means a classical simulator should in principle be able to track only a truncated state vector holding a fraction of the total probability mass. This work implements exactly that: a sparse representation storing only nonzero amplitudes, fully vectorized operations, and optional hardware acceleration, released open-source. It sits at the classical-simulation boundary that quantum advantage claims are measured against.

Susquehanna International Group, one of ByteDance's largest shareholders, is winding down its China-based venture investment team; SIG China's managing director Tim Gong is expected to leave. It is another marker in the steady withdrawal of U.S. venture capital from Chinese technology, which matters for AI specifically because it further decouples the funding bases of the two ecosystems at a moment when Chinese open-weight models — GLM, DeepSeek, Qwen, MiniMax — are the most credible price competition the frontier labs face.

PAST-TIDE reframes stance detection as cloze-style masked language modeling, letting a verbalizer map label words to stance categories through the pretrained MLM head rather than bolting on a randomly initialized classification head. It adds prototypical contrastive learning with learnable class prototypes, which makes contrastive training batch-size independent, and topic-conditional layer normalization for cross-topic transfer. Macro-F1 of 0.75 and 0.74 on the two shared-task subtasks. The general lesson is the transferable one: minimal architectural additions to a pretrained model remain competitive in low-resource settings.

Adam Brown returns to distill general relativity down to the single idea at its heart, aimed at listeners who will never take the twenty-lecture graduate course he taught at Stanford. Not an AI episode, and it does not pretend to be — included because Brown runs Google's BlueShift team and the show is a standing source here. Worth the time on its own terms; it will not change your model of the field.