Wolf Digest — Tuesday, June 9, 2026

A new White House memo directs U.S. national-security agencies to accelerate AI adoption, with Defense One reporting that the FBI and the Office of the Director of National Intelligence are specifically told to build deep, proactive relationships with frontier-AI companies. The directive fits the administration's broader push to remove friction from government AI procurement and integration, and dovetails with the more hands-off regulatory posture that OpenAI cited in its IPO calculus the same day. The move further entangles the intelligence community with a small set of frontier labs, raising familiar questions about vendor concentration, oversight, and the security of model access.

SWE-Explore argues that repository-level benchmarks like SWE-bench collapse coding-agent capability into a binary resolved/unresolved label, hiding the fine-grained skills that actually matter: repository understanding, context retrieval, code localization, and bug diagnosis. The benchmark isolates the exploration step, giving an agent a repository and an issue and scoring how well it navigates to the relevant code before any patch is attempted. It is one of several entries this week pushing coding-agent evaluation toward process-level rather than outcome-level measurement.

SpatialWorld targets a gap in multimodal-LLM evaluation: existing spatial benchmarks rely on passive static VQA or simulator-specific pipelines and do not test general interactive spatial understanding. It unifies eight heterogeneous simulation backends under a shared interface so the same agent can be evaluated on interactive, real-world-style spatial tasks across environments. The work is part of a visible cluster of world-model and spatial-reasoning papers this cycle probing whether MLLMs can actually perceive and act in 3D rather than answer questions about static images.

ResearchClawBench evaluates whether AI coding agents can autonomously reproduce real scientific results. Each of its 40 tasks is grounded in a real published paper, supplies the related literature and raw data, and hides the target paper during evaluation; expert-curated multimodal rubrics decompose the target artifacts into weighted criteria for re-discovery scoring. It joins a wave of benchmarks trying to move agentic-science evaluation past toy problems toward verifiable, paper-level reproduction.

Databricks has discussed a new funding round that could launch within a month and lift its valuation to between 165 and 175 billion dollars, up from 134 billion late last year, according to The Information. The 13-year-old data and AI platform has repeatedly chosen successive private rounds over going public, a posture that looks increasingly distinctive on a day when OpenAI and Anthropic are both filing to IPO. The raise underscores how much private capital remains available to the largest AI-infrastructure companies even as the cost of compute climbs.

This paper diagnoses why strong generative LLMs make poor off-the-shelf embedding models: when text embeddings are projected onto the vocabulary space via the unembedding matrix, they align with frequent but uninformative tokens, and this over-expression of high-frequency tokens suppresses the semantic signal. The authors reframe the unembedding matrix as a feature lens to identify and counteract the effect, improving performance on massive text-embedding benchmarks without retraining the base model.

Artificial Analysis published an evaluation putting MiniMax-M3 at the top of its open-weights tier on the Intelligence Index v4.0, scoring 54.7 and trailing only proprietary frontier models (Claude Opus 4.8 at 61.4, GPT-5.5 at 60.2, Gemini 3.1 Pro at 57.2). The headline caveat is that the result stands once the weights are actually released. If the open weights ship as evaluated, MiniMax-M3 would be the strongest openly available model on this composite, continuing the trend of Chinese labs leading the open-weights frontier.

On-policy distillation is increasingly used to improve LLM reasoning, but its training dynamics are poorly understood. Using parameter-space diagnostics, the authors place on-policy distillation in a relaxed, off-principal regime: relative to supervised fine-tuning it updates fewer weights and avoids principal directions more strongly, while relative to RL with verifiable rewards its updates are less tightly concentrated. The analysis offers a mechanistic vocabulary for why distillation behaves differently from the two dominant post-training recipes.

OmniGameArena addresses three weaknesses in VLM game-agent benchmarks: single first-attempt scores, solo-only play, and no unified protocol across commercial VLMs, open-weight VLMs, and specialized game policies. It ships twelve newly built Unreal Engine 5 games spanning solo, player-versus-player, and cooperative modes under a unified action interface, and reports improvement dynamics rather than one-shot scores so agents that learn within an environment are credited for it.

Video world models that keep 3D consistency usually rely on explicit point-cloud memory built in RGB space, which is expensive (repeated rendering and VAE encoding) and lossy (the round trip through pixels discards latent features). This paper introduces latent spatial memory, a persistent 3D cache that stores scene information directly in the diffusion latent space, avoiding pixel-space round trips while preserving geometric consistency across generated frames.

With TSMC struggling to meet overwhelming demand for leading-edge capacity, several major chip designers including Google and Nvidia are quietly turning to Intel as a backup manufacturer for their most advanced processors, according to The Information. The move is a potential lifeline for Intel's foundry ambitions and a hedge for AI-chip designers against single-supplier concentration risk at the most critical node.

Action-conditioned world models generate multi-segment video from a first frame, text prompt, and camera-action sequence, and their characteristic failure is memory rather than image synthesis: when the camera leaves and returns, the scene or a salient object silently changes. Echo-Memory fixes the action-to-video interface so that competing memory designs can be compared without confounds from backbone, training, retrieval, and evaluation differences, providing a cleaner testbed for the field.

Nvidia and SK Hynix signed a multi-year agreement covering the design and manufacturing of advanced memory chips, including memory for Nvidia's next major platform, Vera Rubin, as AI demand strains global memory supply. Locking in a memory partner years ahead is the latest example of frontier-compute vendors securing scarce upstream capacity, mirroring the same scarcity dynamics that are driving banks to explore compute-price hedging.

Long-context inference is bottlenecked by the KV cache, which grows with context length. The paper argues existing KV-compression methods either degrade quality, are too slow to compress a single long prompt, require the input to fit the target context window, or are incompatible with production inference engines. It scales encoder-decoder compressors that map a long token sequence into a shorter sequence of latent embeddings, aiming for compression that is both cheap and engine-compatible.

This work targets adaptive AI red-teaming, where attacker and defender models are co-trained so each improves against the other. Prior efforts using PPO and DPO worked but reported that GRPO is unstable in this setting. The authors introduce AdvGRPO, which stabilizes GRPO for joint attacker-defender optimization using dense multi-channel rewards and decoupled advantage normalization, yielding both novel attacks and more robust defenders.

MMAE is presented as the first comprehensive testbed for general-purpose, instruction-based audio editing. As interactive editing expands from images and video into audio, the authors argue evaluation infrastructure has lagged, remaining fragmented and limited to narrow subdomains or basic operations. MMAE provides a unified, multitask benchmark to measure instruction-following audio edits across many task types.

Most tool-integrated-reasoning benchmarks evaluate LLM agents only on idealized happy paths, ignoring real-world tool failures. ToolMaze introduces dynamic path discovery and error recovery, using a two-dimensional design: DAG-based topological complexity and a two-by-two taxonomy of tool perturbations (explicit/implicit, transient/permanent). Across models, perturbations degrade performance substantially, exposing how brittle current agents are when tools misbehave rather than simply needing the right call.

Using frozen-feature probing on IntPhys2 and Minimal Video Pairs, this study tests whether pretrained video models encode intuitive-physics information and how it varies by family, layer, and probe. Predictive joint-embedding V-JEPA achieves the strongest results, especially with probes modeling temporal dynamics, outperforming masked-reconstruction VideoMAE and a diffusion-based generator (LTX-Video). The result is evidence that predictive joint-embedding objectives capture physical structure better than reconstruction or generation.

This mechanistic case study asks what values RLHF actually encodes and finds that, for partisan political orientation, alignment training produces functional compliance rather than deep change: the model presents as neutral on the surface while the underlying partisan structure remains intact in its representations. It adds to a growing body of evidence that RLHF yields shallow behavioral masking rather than genuine value alignment, with implications for how much to trust surface-level neutrality.

Progress in genomic foundation models is hard to assess because of fragmented benchmarks and incompatible protocols. GENEB evaluates frozen representations from 40 genomic foundation models across 100 tasks in 13 functional categories under a single probing-based protocol, including few-shot regimes, enabling controlled comparison across model scale and architecture. It is a diagnostic step toward making cross-model claims in genomics actually comparable.

Engineers from Crusoe, a data-center developer for OpenAI and Oracle, are working overtime in Abilene, Texas to get natural-gas turbines to operate harmoniously with one of the most expensive AI supercomputers ever built, and it is proving harder and costlier than expected, per The Information. The site is part of OpenAI's Stargate effort, and the cost and power challenges underscore that the binding constraint on frontier compute is increasingly energy and construction, not chips alone.

Offline safe RL is attractive for safety-critical systems like robotics because it learns from static datasets without online interaction, but that reliance exposes it to data-poisoning attacks that inject malicious samples to induce unsafe behavior. Safe-RULE (safe reinforcement unlearning) is proposed as a defense that removes the influence of poisoned data without retraining the policy from scratch, restoring safety guarantees after an attack is detected.

Two U.S. combatant commanders leading nuclear and mobility forces say they are using AI to generate higher-quality options for logistics, sustainment, and operational maneuvers, telling DefenseScoop they keep finding new areas to apply the tools and reaching solutions faster. The comments are a concrete data point on operational, rather than experimental, military AI use in planning, and arrive alongside the Trump memo pushing national-security agencies to accelerate adoption.

ElevenLabs signed a Memorandum of Understanding with the UK Department for Science, Innovation and Technology covering three areas: voice AI for public-service accessibility (blind and visually impaired users, low literacy, the elderly, and linguistically diverse communities including Welsh), deepened AI-security research with the UK AI Security Institute on how people perceive AI-generated voices and whether they can tell they are speaking to an AI, and talent development. ElevenLabs will triple its London headquarters and double UK headcount to 200. It already runs government voice deployments in Ukraine, the Czech Republic, and Greece.

The Office of Personnel Management added a new batch of industry partners to its Tech Force hiring program as it takes root across agencies: Cisco, Scale AI, Wiz, Arista Networks, Armada, Cognition AI, Cognizant, Payward, and Moveworks, per FedScoop. The inclusion of AI-native firms like Scale AI and Cognition AI signals the federal government's intent to pull commercial AI talent and tooling directly into agency workforces.

Tools for Humanity, Sam Altman's identity-verification company behind the eye-scanning Worldcoin/World project, is reportedly struggling to generate revenue and will downsize its staff, TechCrunch reports, in a notable contrast to OpenAI's IPO momentum the same day. The juxtaposition highlights how uneven the returns are across Altman's constellation of ventures even as his flagship races to the public markets.

Mercor CEO Brendan Foody publicly accused Sequoia, and named it as one of several top firms, of selling the same equity at two different prices, a practice he calls dual-pricing. The dispute is a window into the increasingly aggressive valuation mechanics around AI-data-labeling startups like Mercor, whose bespoke human-expert trajectories underpin the very RL training pipelines Dwarkesh Patel dissected the same day.

DARPA invited its first tranche of teams to compete in the Lift Challenge, which carries 6.5 million dollars in prizes. The program is part of DARPA's continued use of prize competitions to pull novel technical approaches from non-traditional performers into defense-relevant problem spaces.

Italian conglomerate Bending Spoons, which has rolled up aging U.S. internet businesses including AOL, Eventbrite, Evernote, and Vimeo, filed to go public in the U.S. to raise funds for further acquisitions, per The Information. It adds to a crowded 2026 IPO pipeline that now includes OpenAI, Anthropic, and SpaceX, though Bending Spoons's playbook is acquisition-and-optimize rather than frontier AI.