Wolf Digest — Friday, July 24, 2026

A new executive order, 'Securing America's Defense Supply Chains and Ensuring Domestic Acquisition of Critical Materials,' authorizes the Department of Defense to use AI to map vulnerabilities in its supply chains and tightens the use of waivers for acquiring covered materials under 10 U.S.C. 4872. The order directs the Secretary of Defense to develop implementation guidance aimed at reducing reliance on materials sourced from geopolitical adversaries and building resilient domestic and allied supply chains. It is a concrete instance of AI being written directly into procurement and industrial-base policy rather than into weapons systems.

The DoD announced an Enterprise Software Agreement with Oracle worth nearly $7 billion that consolidates the company's software, services and licenses used across the Pentagon, Coast Guard and Intelligence Community into a single contract vehicle. Negotiated by the Department of the Navy with a five-year base and five-year option, the deal provides no new funding but restructures existing contracts; the Pentagon projects at least $441 million in savings. It reflects the department's broader push to rationalize enterprise IT spending as AI and cloud workloads scale.

AREX introduces a family of recursively self-improving deep-research agents that exploit the asymmetry between discovery (expensive) and verification (decomposable into constraint-wise checks). Rather than only searching longer, AREX alternates an inner loop that gathers evidence and builds a provisional answer with an outer loop that audits that answer constraint-by-constraint, flags unresolved claims, and launches targeted follow-up research from the partially-verified state. The framing formalizes something practitioners do informally and turns verification into an explicit optimization signal for multi-constraint research tasks.

On-policy self-distillation normally needs asymmetric information (privileged answers or extra visual evidence) so the self-teacher outsignals the student. VCSD removes both, deriving the asymmetry purely from input conditioning: at each student-generated prefix, an EMA teacher produces two next-token distributions under the same prompt and prefix, one conditioned on the original image and one with the image content removed, and the contrast becomes the distillation signal. It is a notably simpler recipe for teacher-free on-policy self-distillation in vision-language models, and its wide aggregator pickup (six surfacing sources) signals real interest.

Frontier models increasingly ship a built-in Multi-Token-Prediction draft head for speculative decoding on the assumption it is nearly free, but at million-token context that breaks: the MTP head runs full attention over the entire KV cache every draft step, so its read grows linearly with context and dominates draft cost exactly where speculation is most valuable, and a deep native draft can go net-negative (slower than no speculation), worsening under hybrid/linear-attention targets. Windowed-MTP applies a StreamingLLM-style sliding window to the draft's attention to remove that tax. A targeted efficiency fix for long-context serving.

Stripe is in talks to acquire OpenRouter, the router that lets developers access hundreds of AI models through one interface and picks the best or cheapest option per request, for close to $10 billion, according to people familiar with the talks. That is a roughly eightfold markup on OpenRouter's $1.3 billion valuation and follows reports that Databricks also held early discussions. A deal could be announced within a month or could still fall apart. The price signals how strategically valuable the model-routing layer, the neutral switchboard sitting between applications and a fragmenting field of frontier and open models, has become as inference commoditizes.

GraphVid conditions image-to-video generation on structured interaction graphs, letting users specify precise multi-object interactions that text prompts and trajectory scribbles handle poorly, especially under occlusion or overlap. The authors curate GraphVid-Bench, a large interaction-centric video dataset with relational annotations, and report strong control while using substantially less supervision than trajectory-based methods. It advances controllable video generation from pixel-motion constraints toward explicit relational structure.

GS-Agent builds dynamic, physically-realistic 4D worlds from natural-language descriptions by wrapping foundation models in an agentic system that emulates how humans author 4D content, but automated end-to-end, with physics engines kept in the loop to enforce plausibility and controllability. It stakes out a middle path between manual computer-graphics authoring and purely generative 4D methods that struggle with physical consistency, relevant to simulation, robotics data and world models.

The Office of Naval Research's new science-and-technology strategy leans more heavily on AI as a tool for doing its own research and for speeding delivery of technology to the fleet, spanning a portfolio that includes AI and autonomy, directed energy, C5ISR and naval space, and undersea systems. The Chief of Naval Research framed strategy as placing informed bets and continuously reallocating the portfolio to its highest and best use. It is an institutional signal that AI is being treated as research infrastructure inside a major defense S&T funder, not just a program area.

An analysis argues the Pentagon's next advantage in fires will come from weapons that adapt in flight, share data across the kill chain, scale affordably and earn enough trust to deploy without a human beside them, tying together the Air Force's Family of Affordable Mass Missiles and the push to field 10,000 low-cost cruise missiles within three years. The 'adaptive lethality' framing captures the shift from precision alone toward volume, speed and software-mediated autonomy in munitions, the demand signal behind much of the defense-AI contracting activity.

OpenForgeRL is an open-source framework for training agents that live inside elaborate inference harnesses like Claude Code, Codex and OpenClaw end-to-end, which existing SFT/RL stacks cannot natively express because those harnesses are stateful and multi-process. It inserts a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase such as veRL, and a Kubernetes orchestrator that runs each rollout in its own remote container, enabling scalable training on any harness in any environment. It targets a real gap between how capable agents are built and how open infrastructure can train them.

This paper shows that adding a dense next-observation-prediction reward to sparse-reward long-horizon LLM agents does not just fail under group-normalized RL (GRPO), it destroys the policy. Across Qwen3 at 1.7B/4B/8B on ALFWorld, the potential-based prediction reward drives every run into a degenerate absorbing state, prediction accuracy going to 1.0 while task success collapses to 0 and episodes pin at the horizon, the 'dark room' pathology. A single-factor ablation localizes the cause to GRPO's standard-deviation normalization: removing only that turns the same reward from catastrophic to baseline parity, with a two-line argument for why all-fail groups make the z-scored advantage invariant to reward shaping. A crisp, mechanistic warning for anyone bolting dense rewards onto GRPO agents.

WorldWeaver augments autoregressive video-diffusion rollouts with learnable 'world-state registers': tokens that store shared world information, track per-agent status and update after each generated chunk, so multi-agent, multi-view interactive world models can maintain consistent shared state instead of only carrying forward observation history as conditioning. Registers are grounded with supervision spanning per-agent status, global bird's-eye views and scene text, improving persistence of state across agents and views.

Self-Forcing trains video-diffusion students on their own rollouts to cut exposure bias, but the historical KV cache is used by future frames only as frozen state, so future losses cannot supervise how earlier latents are written into keys and values, the 'historical context-gradient gap.' Self Gradient Forcing restores that signal with a two-pass strategy: a no-gradient autoregressive rollout matching inference, then a gradient pass from a sampled denoising exit step, without backpropagating through the full serial rollout. It targets long-video extrapolation quality.

VLMs often solve a geometry problem from its text view but fail on the equivalent diagram, or vice versa, revealing that different views expose complementary reasoning paths and failure modes. MIRROR builds ODA-Data, paired text-dominant, image-dominant and combined views of the same problems, and trains models to learn from the other view, exploiting the inconsistency that standard multimodal post-training leaves on the table. It is a concrete lever for the persistent weakness of visual reasoning in VLMs.

Real-world agent learning is bottlenecked by costly environment interaction. This work formalizes 'experience distillation': internalizing an agent's own interaction histories into model weights via context distillation, so the sample-efficiency gains of in-context learning survive after the experience is removed from the prompt, without any further environment interaction beyond the already-collected experience. Experiments span 749 curated software-engineering tasks and six text-adventure environments, targeting the gap where in-context learning helps but its benefits vanish once context is dropped.

The paper argues PPO-Clip's exploration collapse in LLM RL stems from a geometric flaw: it implicitly measures policy discrepancy with a Euclidean metric that is inconsistent with the intrinsic geometry of the policy manifold, yielding updates that are too conservative in low-probability regions and too aggressive in high-probability ones. Riemannian Isometric Policy Optimization (RIPO) enforces isometric updates on the policy Riemannian manifold to correct this, offering a principled account of a widely-observed failure mode rather than another heuristic clip variant.

VLM-IE3D equips vision-language models with both Implicit Geometry Tokens that capture high-level geometric priors and Explicit Geometry Tokens that encode reconstructed 3D structure, fused with 2D visual cues through a 3D-aware adapter, all learned from RGB videos with no depth sensor. The RGB-only design aims to give VLMs the fine-grained spatial understanding they typically lack on 3D tasks, a recurring bottleneck for embodied and spatial applications.

Reasoning models such as DeepSeek-R1-Distill-Qwen-7B show a bimodal pattern: generations either converge within the token budget (90.3% accuracy on AIME 1983-2024) or exhaust it without concluding (6.6% accuracy), with a 62% overall convergence rate. Linear probes on hidden states detect the outcome early, layer-20 activations at token 150 reach AUC 0.608 and stay above chance even at token 50, outperforming behavioral baselines. It points toward cheap early-exit or budget-reallocation policies for test-time compute.

MAPS (Master-Agent Proto-plan System) tackles multi-agent coordination at unsignalized intersections with a hierarchical DRL design: a centralized Master agent emits a compact continuous embedding, a 'proto-plan,' encoding a global coordination strategy, while decentralized Worker agents combine it with local observations for vehicle-specific control. Decoupling strategic intent from tactical execution sidesteps combinatorial action spaces and reliance on privileged information, addressing a persistent MARL pain point for autonomous driving.

OpenAI President Greg Brockman said he supports a suggestion from Elon Musk that leading AI developers meet every few weeks to discuss and share safety and security concerns, calling it a reasonable baseline proposal. The endorsement is notable given the Musk-OpenAI history and lands amid the broader fight over frontier-model governance and Chinese open-weight restrictions, where the major labs are trying to shape the coordination mechanisms before regulators impose their own.

Intel shares rose 12% after-hours after reporting its fastest revenue growth in 15 years, driven by AI-customer chip demand; its data-center and AI hardware unit grew 59% to $6.3 billion, and the company generated $4.5 billion in cash. After years of being written off in the AI compute story, the print suggests Intel's CPU franchise and data-center business are getting real pull from AI buildouts.

Google is closing in on a billion monthly users for Gemini, up from over 750 million in February, putting the assistant on track to join Google's roster of billion-user products. Distribution through Search, Android and Workspace is the engine, and the scale matters competitively: it gives Google a consumer usage base that pure-play labs cannot match and a feedback loop for improving models.

Latent reasoning carries intermediate computation as continuous vectors and can match explicit chain-of-thought at far shorter horizons, but latent reasoners have stayed imitation-bound because they lack a tractable per-step likelihood and an adaptive stopping interface, so outcome-reward RL cannot elicit latent test-time scaling. Surrogate Latent Policy Optimization (SLPO) introduces a surrogate policy that gives latent trajectories the missing likelihood and stopping controls, bringing outcome-reward RL to latent reasoning and pushing it past pure imitation.

This work studies hypernetworks for train-time knowledge injection: given a large corpus of facts, train a hypernetwork to emit a fixed LoRA adapter that, inserted into a target model, lets it answer questions about those facts. The design decouples the hypernetwork's injection capacity from the target's general capability and characterizes how the ability scales, an underexplored regime, offering a route to reliable large-scale factual editing beyond per-fact fine-tuning.

Behavior-cloning fine-tuning of a VLM into a vision-language-action policy progressively overwrites the pretrained representations that support visual and semantic generalization, and web-data co-training does not fix the resulting language-action misalignment. Anchor-Align adds two objectives: Vision-Language Anchoring distills layer-wise features from a frozen VLM copy to prevent drift, and Language-Action Alignment converts each action target into a discrete motion-direction label jointly supervised with language. It directly targets the generalization loss that standard manipulation benchmarks fail to expose.

Governance rules often forbid sending sensitive research data to third-party cloud LLMs, so this work introduces an open-source framework to evaluate locally-deployable open-weight models as agents on data preparation for longitudinal population studies, a persistent research bottleneck. It ships a curated ground-truth dataset (cleaning scripts across six sweeps of a British cohort study) and task definitions, quantifying how far local open-weight agents can go on privacy-constrained data-wrangling where cloud models are off-limits.

This essay documents that LLMs systematically overuse epanorthosis, the 'this is not X, it is Y' self-correcting figure Cicero and Quintilian catalogued, and argues it is a trained disposition driven mainly by a training distribution rich in promotional prose and by preference tuning that rewards confident, emphatic phrasing, with left-to-right generation an amplifier rather than the root cause. It proposes an Epanorthosis Index scoring the figure against genre-specific human baselines and a mitigation program, a stylometric lens on how RLHF shapes model prose.

Amazon closed its AGI Lab, a San Francisco research-and-product team developing software to improve the usefulness of AI agents, as part of layoffs in its artificial-general-intelligence unit. The shutdown of a dedicated agent-research group at a hyperscaler, even as agents dominate the industry narrative, is a signal about how Amazon is concentrating its AI bets and where it judges internal research to be duplicative.

AMD detailed Helios, a rack-scale system that packages its Instinct accelerators, CPUs and networking into an integrated unit meant to compete directly with Nvidia's rack-scale platforms, with shipments starting later this year. Rack-scale integration, selling the whole rack as the unit of compute rather than individual GPUs, is now the battleground for large-scale training and inference deployments, and Helios is AMD's clearest move onto that turf.

This EEG foundation model combines a Mamba-based raw-signal encoder, a ViT-style encoder for time-frequency representations and a lightweight text encoder in a shared embedding space, pretrained with masked modeling, cross-view contrastive alignment and temporal-consistency losses. The goal is an adaptable EEG backbone that transfers across datasets and tasks rather than the dataset-specific models common in epilepsy work, a template for foundation models on physiological time series.

Real interaction is dynamic: users disclose, revise and redirect intent mid-conversation, yet LLMs are still mostly trained and evaluated in single-turn, fully-specified settings. This paper introduces a framework that turns static single-turn tasks into dynamic multi-turn conversations with evolving intent while preserving each task's ground truth, then shows how poorly current models track and act on that shifting intent, exposing a blind spot for collaborative-agent deployments.

AMD said it is working with Cerebras, a rival AI-server-chip developer, to connect AMD's server racks to Cerebras wafers so both companies' chips can run the same workload simultaneously, an example of disaggregated inference, splitting a single model's serving across heterogeneous accelerators. The tie-up between competitors underlines how inference economics are pushing the industry toward mixing chip types to optimize cost and latency rather than standardizing on one vendor.

Runway launched Media Router, a tool that automatically picks the best image, video or audio model for a given request based on whether the developer prioritizes quality, speed or cost. It brings the model-routing pattern, already common for text LLMs, into generative media, an implicit acknowledgement that no single media model dominates and that orchestration across many is now the product.

The Selective Service System is soliciting software to modernize and speed the administration of a potential military draft, including registration and processing systems. The procurement is a logistics-and-IT modernization effort rather than an AI-capability story, but it is a notable data point in the wider trend of national-security agencies digitizing core functions.

Etched, founded by three Harvard dropouts, raised from big-name investors at a $10.3 billion valuation on the pitch that its transformer-specialized chips and memory components accelerate inference on any model without GPUs. The raise is a bet that fixed-function silicon tuned to the transformer can beat general-purpose GPUs on inference cost, and investors are underwriting it despite skepticism about single-architecture ASICs in a fast-moving model landscape.

Anthropic upgraded Claude's voice mode with more capable underlying models and the ability to take actions such as rescheduling a meeting or drafting an email during a spoken conversation. The update lands the same day OpenAI shipped a consumer ChatGPT Voice refresh and an enterprise Presence product, marking voice as an actively contested surface among the frontier labs.

ElevenLabs added References to its Music v2 model, letting users upload a 10-second-to-5-minute track so the model matches its style, instrumentation and feel, optionally combined with a text prompt. Every reference passes a copyright check that blocks tracks matching recordings owned by others, and the feature is exposed in ElevenMusic, ElevenCreative and via the API. It is an audio analogue to image style-reference conditioning, trading a text description for a concrete acoustic anchor.

OpenAI opened ChatGPT Health to all U.S. users, letting people integrate personal data from services like Apple Health, Function and MyFitnessPal to get health-related answers grounded in their own metrics. Broad consumer health features raise the usual questions about accuracy, privacy and the boundary with medical advice, but they also mark OpenAI's push to make ChatGPT a daily utility around personal data.

AegisAI, founded by former Google security executives, raised $36 million to fight AI-generated spear phishing using agents that analyze each message the way a human would, catching small anomalies that rule-based filters miss. The funding reflects how quickly generative models have raised the quality of targeted phishing and the emergence of a defensive-AI market to counter it.

Offensive-security researchers who hunt vulnerabilities and build exploit tooling told TechCrunch that OpenAI's and Anthropic's safety guardrails increasingly get in the way of legitimate work, refusing tasks that are indistinguishable, to the model, from malicious ones. The tension illustrates the dual-use problem at the heart of model safety policy: the same refusals that block abuse also block sanctioned red-teaming and defensive research.