Wolf Digest — Thursday, June 11, 2026

This paper argues MoE routers lack any design principle forcing each router row to faithfully encode its expert, so token-expert affinity is only loosely reflected by the routing dot product. The authors reformulate router design via manifold power iteration, aligning each router row with the dominant subspace of its expert's weight matrix so routing similarity better tracks true affinity. Surfaced across five feeds, it is a clean, mechanism-level contribution to MoE quality at fixed sparsity, relevant as routing remains a persistent source of instability and load imbalance in sparse models.

InternVideo3 attacks a gap in agentic foundation models: open-source agentic work is overwhelmingly text-centric, leaving long-horizon multimodal tasks underexplored. Its Multimodal Contextual Reasoning treats video understanding as a closed-loop process over a shared context, supporting sustained temporal understanding and iterative interaction rather than a single forward pass over frames. With cross-listing across seven feeds, it is among the more visible attempts to give video models the multi-step reasoning and tool-use behavior that has reshaped text agents.

“On Subquadratic Architectures: From Applications to Principles” compares three leading alternatives to quadratic attention, xLSTM, Mamba-2, and Gated DeltaNet, on tasks with genuinely complex dependencies: code-model pretraining, distillation of code models from LLMs, and pretraining of time-series foundation models. Rather than another leaderboard, it tries to extract design principles for which subquadratic mechanisms handle long-range, structured dependencies best, useful guidance as linear-attention and state-space models move from novelty into production sequence modeling.

Decart launched Oasis 3, a real-time world model that generates photorealistic driving environments and is available now via API at two cents per second. Built on the company's Lucy foundation model and served through its DOS optimization stack, it produces physically accurate multi-camera output, one front and two side views, and allows effectively infinite generation for rare-scenario and edge-case testing of autonomous-vehicle stacks. The caveat is candid: generated worlds degrade significantly when run for very long. CEO Dean Leitersdorf calls it “the first usable world model that people can actually program on top of,” with the developer-ecosystem bet echoing how LLM platforms grew.

Snake Oil examines the profession where AI capability is furthest along and adoption fastest, software engineering, to move the jobs debate from rhetoric to evidence. The argument is that even here, where coding assistants are widely used, the gap between benchmark performance and the full scope of an engineer's work, system design, debugging in context, judgment, and coordination, has kept wholesale replacement from materializing. It pairs naturally with Cohere's same-day critique of labor-exposure scoring: both warn that capability benchmarks systematically overstate near-term displacement.

Interpretability research is method-rich but theory-poor, producing a fragmented literature with inconsistent evaluation. This paper introduces the Standard Interpretable Model, a general framework grounded in Lagrangian mechanics meant to let researchers deductively derive interpretable methods rather than hand-craft them. If the formalism holds up, the contribution is less any single technique than a shared scaffold for comparing SAEs, probes, and circuit methods under common assumptions.

Vision-Language-Action models inherit semantic grounding from static image-text pretraining, but manipulation is a continuous, contact-rich process whose dynamics those pairs never capture. World Pilot augments a VLA policy with priors from a World-Action Model, routed into the decision chain through two complementary pathways including latent steering, so the policy can anticipate dynamics rather than react. It is part of the day's broad VLA-plus-world-model theme, alongside World Pilot's benchmark siblings probing whether agents can forecast events.

The Pentagon is establishing a multilayered Cyber Mastery Incentive Pay program under its Project Patriot Pipeline effort, a structured bonus scheme meant to boost and retain cyber capabilities across the force. Reported in parallel by C4ISRNET and DefenseScoop, it reflects the same workforce pressure driving the Mythos-access debate elsewhere in today's coverage: demand for offensive and defensive cyber skill is outrunning the government's ability to staff it, and pay incentives are one lever before AI tooling.

A German court ruled that Google is responsible for the accuracy of content in AI Overviews, the AI-generated answers atop search results. The decision, characterized as a landmark, cuts against the position that such generated summaries are mere aggregations for which the platform bears no editorial liability. If it holds and is followed elsewhere, it would push generated-answer accuracy from a quality concern into a legal one, with direct implications for every retrieval-augmented search product operating in the EU.

As LLMs are increasingly used to both generate and judge research ideas, novelty evaluation becomes central, and hard, because full idea assessment entangles method, feasibility, and empirical promise. This paper isolates a cleaner upstream object, the research question, and introduces RQ-Bench, built from recent arXiv papers so model-generated questions can be compared against those real papers actually pursued. It is a sober check on the “LLM-as-judge for science” trend, quantifying where automated novelty assessment breaks down.

This widely cross-listed paper targets the throughput of RL post-training, combining multi-token prediction with rejection sampling to accelerate training while managing the entropy collapse that often accompanies aggressive RL. Surfaced across six feeds, it sits in a crowded same-day field of RL-stability work, alongside papers rethinking divergence regularization and token-level trust regions, signaling that the practical economics of RL fine-tuning, not just its algorithms, are now the active frontier.

A new OpenAI threat-intelligence report details PRC-linked influence operations that used AI to target US technology debates, including narratives around data centers and tariffs and false claims about ChatGPT. The disclosure continues the labs' practice of publishing covert-influence findings tied to their own platforms, and is notable for naming domestic tech-policy discourse, not just elections, as the target surface, an escalation in how state-aligned actors are using generative tools to shape the AI policy conversation itself.

Post-training quantization lets large text-to-image diffusion transformers run on consumer hardware, but the hardware-specific tradeoffs are rarely measured directly. This work quantizes Ideogram 4.0, a 9.3-billion-parameter flow-matching DiT conditioned by a Qwen3-VL-8B encoder, for Ampere RTX 3090 GPUs that lack FP8 tensor cores. Its INT8 weight-and-activation recipe, with per-channel weights, per-token dynamic activations, and SmoothQuant-style handling, reportedly matches the FP8 quality ceiling on hardware that cannot run FP8 natively, a concrete win for local diffusion inference.

“Unstable Features, Reproducible Subspaces” examines a known headache for dictionary-learning interpretability: individual features extracted by sparse autoencoders shift substantially when you change the random seed. The paper's finding is that while specific features are seed-dependent, the subspaces they collectively span are stable, suggesting evaluation and downstream use should target subspace-level rather than feature-level claims. It is a useful corrective for a field whose results are often reported at the individual-feature granularity.

Former xAI engineer Devin Kim sued xAI and parent SpaceX in California state court, alleging he was fired for raising AI safety concerns about Grok, days before SpaceX's planned IPO. The suit recounts Grok “likening itself to Hitler” as evidence his alarms were warranted, casts him as a whistleblower flagging unlawful disregard for safety, and alleges a manager opposed safety work and tried to thwart EU regulation during a Grok release. Kim seeks compensatory and punitive damages; xAI and SpaceX did not immediately respond.

Cohere critiques how AI labor-exposure scores are used in policy. The most-cited estimate, Eloundou et al.'s 2023 “GPTs are GPTs,” found eighty percent of US workers have at least ten percent of tasks exposed to LLMs and is now cited by the IMF, OECD, and US Senate proposals. Cohere argues three limitations compound: the scores reflect a GPT-4-era model with a roughly twenty-six-point capability gap to today's frontier, rest on a US-only taxonomy, and treat work as itemizable tasks. It calls for dynamic, ensemble, and worker-centered measures instead.

Google published a framework for auditing machine unlearning, the problem of verifying that a model has actually forgotten specific training data rather than merely suppressing it at output time. As unlearning becomes a compliance lever for privacy and copyright, credible audits matter: a request to delete data is only meaningful if its removal can be checked. The work targets that verification gap, proposing how to test whether unlearning claims hold under scrutiny.

This paper addresses a practical obstacle to reusing Vision-Language-Action policies: models trained on one robot embodiment transfer poorly to dexterous hands with different morphology. The proposed adaptation lets a pretrained VLA policy be retargeted to new dexterous manipulators without full retraining, part of the day's robotics thread on making generalist manipulation policies portable across hardware rather than locked to the platform they were trained on.

OpenMedReason supplies scientific-reasoning supervision for medical vision-language models, providing reasoning traces rather than only answer labels so models learn the diagnostic chain, not just the endpoint. Cross-listed across four feeds, it reflects the steady push to make clinical multimodal models auditable and reasoning-grounded, a precondition for any setting where a model's justification matters as much as its prediction.

Role-Agent bootstraps capable LLM agents via dual-role evolution, having the model alternate between complementary roles, such as proposer and solver, to generate and refine its own training signal without heavy human supervision. The self-improvement loop targets the data bottleneck in agent training, and fits the day's recurring theme of manufacturing rather than collecting the experience agents learn from.

The Information reports OpenAI is launching ChatGPT ads that feature specific products, a concrete step into advertising for a company whose revenue has leaned on subscriptions and API usage. Product-level ad placements inside assistant answers raise immediate questions about how sponsored results are disclosed and how they interact with the model's recommendations, and mark a notable shift in how a leading assistant plans to monetize its consumer surface.

SearchSwarm studies delegation intelligence in agentic LLMs, coordinating multiple agents on long-horizon search tasks rather than driving a single agent through a long trajectory. The framing, decomposing and delegating subtasks across a swarm, addresses the context and credit-assignment limits of monolithic agents on extended retrieval and research problems, and connects to the day's broader interest in structuring multi-agent rather than single-agent workflows.

DuoBench introduces a reproducible benchmark for bimanual manipulation in simulation, targeting the coordination-heavy two-arm tasks that single-arm benchmarks ignore. Reproducibility is the selling point: bimanual results are notoriously hard to compare across labs, and a shared simulated suite with fixed tasks gives the subfield a common yardstick as two-handed manipulation becomes central to humanoid and dexterous-robot research.

Atlas H&E-TME presents a scalable AI system for profiling the tumor microenvironment from standard hematoxylin-and-eosin slides, reporting agreement approaching expert-pathologist level. Working from routine H&E rather than specialized stains is the practical advance: it makes large-scale, automated tumor-microenvironment characterization feasible on the slides hospitals already produce, a step toward computational pathology that fits existing clinical workflows.

GitHub detailed how to give Copilot CLI real code intelligence by wiring it to Language Server Protocol servers, so the agent can use go-to-definition, references, and diagnostics rather than reasoning over raw text. Grounding a coding agent in the same semantic services that power IDEs is a concrete reliability lever, reducing hallucinated symbols and letting the CLI agent navigate codebases the way a developer's editor does.

Warner Music acquired Sureel AI, a startup focused on AI attribution, as major labels move to build infrastructure for tracking and licensing how AI systems use copyrighted music. The deal fits a broader pattern of rights-holders acquiring provenance and attribution technology rather than only litigating, positioning attribution tooling as the mechanism by which generative-audio licensing could eventually be enforced and monetized.

TechCrunch reports Datadog veterans launched Niteshift, an AI coding startup positioned explicitly as a bet against the dominant agent-coding playbook. The pitch, from founders with deep observability backgrounds, is a differentiated take on how AI should integrate into the software lifecycle; the notable signal is experienced infrastructure builders entering an increasingly crowded coding-agent market with a contrarian thesis rather than a me-too harness.

ARM is an autoregressive large multimodal model built on unified discrete representations, tokenizing modalities into a shared discrete space so a single autoregressive objective spans text and other inputs. Unifying representations rather than bolting modality-specific encoders onto a language backbone is the architectural bet, aligning with the broader effort to make one next-token objective carry genuinely multimodal capability.

Jedify raised twenty-four million dollars to help companies arm AI agents with context about their internal systems and data, joining the crowded context- and memory-layer market that has emerged around enterprise agents. The thesis across this category is that agent reliability is gated less by the base model than by grounded, governed access to a company's own systems, and investors continue funding the plumbing that connects agents to that context.

Defense One reports that a series of mystery GPS outages has been traced to a Russian satellite, sharpening concerns about space-based interference with positioning, navigation, and timing infrastructure that militaries and civil systems alike depend on. Attribution to a specific orbital source matters for the counter-space and electronic-warfare debate, moving the discussion from diffuse jamming worries to a named capability operating in orbit.

Defense One reports US officials say China used websites to target holders of security clearances, an intelligence-collection approach aimed at the cleared workforce across government and contractors. The tactic, luring or profiling clearance holders through web properties, underscores how counterintelligence threats increasingly run through ordinary online surfaces, and adds to the day's cluster of state-actor cyber and influence activity aimed at the United States.