Wolf Digest — Thursday, June 25, 2026

Nine-year-old AI chip startup SambaNova is set to raise between 800 million and 1 billion dollars, pushing its valuation to roughly 10 billion — about five times its prior mark — according to executive chairman Lip-Bu Tan, who is also chief executive of Intel. The raise is the latest sign of how much capital is chasing credible alternatives to Nvidia for inference: developers and cloud providers want lower-cost options, and chip firms whose parts can plausibly compete on serving workloads are commanding sharply higher prices. SambaNova's reconfigurable dataflow architecture targets exactly that inference-serving niche.

On-policy self-distillation, where a single model serves as both teacher and student with the teacher conditioned on a correct demonstration to give dense token-level feedback, reliably lifts pass@1. This paper shows it comes at a hidden cost: rollout diversity shrinks and pass@k curves flatten, so the model gets more reliable on its single best guess while losing the spread of distinct solutions that matters for search, reranking, and reinforcement learning that samples many candidates. The finding is a useful corrective for teams optimizing pass@1 in isolation — the metric improves while an unmeasured capability that downstream RL depends on silently degrades.

As agents are granted tools, APIs, and infrastructure, they become active principals in those systems — and the dominant way to control them places the guardrails inside the agent's own runtime, via system prompts, output filters, and guardrail libraries. The paper's argument is that any control living in the agent's address space is, in principle, reachable and defeatable by the agent or by an attacker who compromises it. It proposes an execution-time safety kernel that sits outside that boundary and mediates the agent's actions at the point of execution, so the enforcement layer cannot be edited away by the thing it constrains. The framing borrows from operating-system privilege separation and is a cleaner statement of why prompt-level and output-level safety are structurally insufficient for agents with real-world reach.

Reinforcement learning is the obvious lever for improving multi-step tool use in LLMs, but in practice it often produces instability or limited gains. The authors document cases where models undergoing RL on tool-use tasks exhibit catastrophic collapse, then trace the failure to weak or misallocated learning signal across long action sequences and show how injecting an explicit supervisory signal stabilizes training and recovers gains. The contribution is diagnostic as much as methodological — it explains a failure mode many teams have hit when pushing agentic RL beyond short horizons, and prescribes where the corrective supervision needs to enter.

Space Systems Command announced that Boeing will provide two additional satellites for the Mobile User Objective System, the narrowband tactical SATCOM constellation that United States forces rely on for beyond-line-of-sight voice and data. The roughly 2 billion dollar award, under the MUOS Service Life Extension program, extends operations of one of the Defense Department's critical communications networks and went to Boeing over Lockheed Martin. The item sits in the digest as applied AI and autonomy increasingly ride on assured military communications; the contract itself is a conventional space-hardware procurement reported here for its place in the defense-infrastructure backdrop.

A central safety goal is determining whether a model is actually misaligned, but most work detects concerning behavior and stops there. This paper's point is that behavior alone does not establish misalignment — a worrying action can arise from benign causes such as confusion or a misread of the situation. It proposes model forensics: investigating the internal causes behind a concerning action to distinguish genuine misalignment from innocuous explanations, rather than inferring intent from outputs. The reframing matters because alignment evaluations that treat any bad output as evidence of misalignment will over-attribute hostility and mis-prioritize which behaviors actually warrant intervention.

Process reward models give step-level feedback, but building them for agentic settings is hard: long horizons, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo value estimation infeasible at scale. The paper argues there is a neglected free lunch in defining a progress advantage — a step-level signal derived from how much an action moves the agent toward task completion — that can be extracted from existing rollouts without new labels. It positions this as a cheap, annotation-free path to dense credit assignment for LLM agents, the bottleneck that has made step-level reward modeling impractical for long-horizon tasks.

Agility Robotics, the humanoid startup that spun out of Oregon State University in 2015 and makes the warehouse-oriented Digit robot, plans to go public via a SPAC in a deal valued at roughly 2.5 billion dollars, expecting to generate about 620 million in proceeds. The move makes Agility one of the first pure-play humanoid companies to seek public markets, and the proceeds are aimed at scaling manufacturing and deployment. A SPAC route, rather than a traditional IPO, is notable for a hardware company still early in commercial revenue, and the valuation will be a public reference point for a humanoid sector that has so far been priced almost entirely in private rounds.

Microsoft Research introduced Talos, a system that automates the iterative reanalysis of genomic data to find diagnoses that an initial pass missed. Rare-disease cases often stay unsolved not because the causal variant is absent from the sequencing data but because knowledge moves on after the first analysis — new gene-disease associations are published, and re-examining old cases against current knowledge is labor-intensive and therefore rarely done. Talos targets exactly that gap, framing reanalysis as a repeatable automated loop rather than a one-time manual effort, which is where a meaningful share of diagnostic yield in undiagnosed cohorts has been shown to hide.

In its first earnings report since going public, AI chipmaker Cerebras forecast a narrower gross margin in its core business, and the stock fell sharply as investors reacted. The chief executive said the margin outlook had been misunderstood, but the sell-off underscores how exposed the new class of public AI-hardware companies is to expectations: wafer-scale economics and large customer concentration make margin guidance the number the market watches. It is a counterpoint to the week's exuberant private-market chip valuations, a reminder that public investors are pricing the same inference-compute thesis with considerably less patience.

HiReLC is a hierarchical, ensemble-reinforcement-learning framework that jointly searches over quantization and structured pruning for deep networks. It decomposes the compression decision across two levels: low-level agents operate independently per block to set local quantization and pruning choices, while a higher level coordinates them toward a global compute or accuracy target. Casting joint compression as a hierarchical RL search is the contribution — most prior work treats quantization and pruning separately or with hand-tuned heuristics, and the hierarchy is what keeps the combinatorial per-block search space tractable.

This paper documents a striking training-dynamics phenomenon. Partway through an ordinary pretraining run, a small language model learns a pronoun-gender rule — cued with a girl's name, it resolves the next pronoun to 'she,' generalizing to held-out probes and reaching 0.94 by step 925. Yet by step 3,500 the same model scores near zero on the same probes: a capability that was acquired and generalizing has been un-learned, the mirror image of grokking. The authors study this 'natural ungrokking' and the asymmetric control over which rules survive pretraining, a result with implications for how we reason about capability emergence and disappearance over a training run rather than just at its end.

Roughly two-billion-parameter GUI-grounding agents are attractive for on-device deployment, accessibility tooling, and cheap iteration, but at that scale two recipe questions are open: how to get bounding-box training data without expensive human labeling, and how to combine supervised fine-tuning with self-distillation effectively. WinDOM proposes a self-family-distillation recipe that generates its own grounding supervision and folds it back into training, targeting the small-model regime specifically. The work is a practical contribution to making competent screen-understanding agents that fit on a phone rather than requiring a server-class model behind an API.

Computer-use agents convert vision-language-model predictions into executable GUI clicks, so reliable uncertainty estimates are essential for knowing when to reject an action, how to calibrate confidence, how to rank miss severity, and where to draw spatial safety regions on screen. Yet evidence on post-hoc uncertainty quantification for these agents has been fragmentary. This paper introduces a benchmark spanning vision-language models and UQ methods to measure how well current approaches estimate when a GUI action is likely wrong — a prerequisite for deploying screen-driving agents with action-level safety gating rather than blind execution.

Sparse Mixture-of-Experts architectures balance parameter scalability against compute, but low-resource languages, starved of high-quality training data, tend to be served poorly by the learned routing. SARA proposes semantically-anchored routing that deliberately directs low-resource-language tokens to experts holding the relevant knowledge, rather than letting frequency-driven routing strand them. The aim is to unlock multilingual capability already latent in a MoE's parameters without retraining from scratch — a targeted fix to a known equity gap in sparse models, where the experts exist but the gating fails to reach them for rare languages.

LLMs reach high accuracy on mathematical reasoning, but separate traces on the same problem diverge — some land on the right answer, others fail. Prior analysis looks at the step, chunk, or sentence level, or at tokens where failure has already happened. This paper isolates 'cliff tokens': individual tokens that act as failure triggers, the point where an otherwise-viable trace tips into an unrecoverable error. Locating failure at single-token resolution, and before the error has manifested, is the novelty, and it suggests intervention points for decoding-time steering or targeted training rather than coarse trace-level fixes.

Subject-driven text-to-video has to satisfy two competing demands: in-domain generation that preserves a reference subject's features as faithfully as possible, and cross-domain generation that keeps the subject's intrinsic identity while moving it into a different visual style. DomainShuttle targets open-domain operation across both regimes, aiming to retain subject consistency without locking the output to the reference domain. The contribution is handling the freeform, open-domain case rather than the narrow single-subject setups that earlier methods assumed, which is the harder and more useful version of the problem for real generation workflows.

Standard multimodal benchmarks score each item on one canonical ordering of options, and so miss whether an order-irrelevant shuffle flips the answer — a baseline reliability property that emerging evaluation guidelines explicitly call for. Facet-Probe is a five-facet audit that measures this order sensitivity systematically across multimodal LLMs. The result is a reminder that leaderboard accuracy can mask brittleness: a model that scores well on the canonical ordering but changes its answer under a benign reshuffle is not actually reasoning over the evidence, and single-ordering benchmarks will not catch it.

Group-based RL post-trains LLM agents on long-horizon, sparse-reward tasks by deriving step-level credit from trajectory outcomes, but this ties a step's credit to its rollout's final outcome — so two semantically near-identical intermediate steps can receive opposite credit purely because their trajectories ended differently. Semantic Consistency Policy Optimization addresses that noise by enforcing that semantically similar steps get consistent credit, decoupling a step's value from the luck of its rollout's ending. It is a targeted fix to a well-known variance problem in outcome-supervised agentic RL.

Vision-language-action models are capped by the imitation ceiling of sub-optimal demonstration data; RL fine-tuning can pass that ceiling but is notoriously sample-inefficient, in part because of catastrophic initial unlearning when RL starts from an imitation-pretrained policy. FORCE introduces a value-calibrated warm-up that stabilizes the transition into RL so the policy improves without first destroying what imitation taught it. The target is the practical pain point that has kept RL fine-tuning of VLAs out of reach for most robotics teams: the enormous interaction budgets required to see gains.

Most vision-language-action models attach an action module to a vision-language backbone and train the whole policy jointly, inheriting strong visual and linguistic priors but leaving the action module to learn physical motion almost from scratch — which is data-hungry and embodiment-specific. This work learns explicit action priors that transfer across embodiments, so motion knowledge gathered on one robot body informs another rather than restarting. Cross-embodiment transfer of the action component, not just the perceptual front end, is the contribution, and it speaks directly to the data-scarcity bottleneck that makes per-robot VLA training expensive.

Medical vision-language models usually emit a diagnosis in a single pass without indicating which image regions support it, which limits clinical utility: the output cannot be audited, and the model may hallucinate findings on a normal scan. This paper adds an ROI-rethink step and synthetic-data augmentation so a brain-MRI model both detects anomalies and points to the regions that justify its reasoning. Spatial grounding is the contribution — tying conclusions to evidence regions is what makes such systems checkable by a radiologist rather than opaque, and it directly attacks the false-positive-on-normal-scan failure mode.

Google Research argues that chain-of-thought reasoning does more than help models work through a problem — it unlocks parametric knowledge the model already holds in its weights but fails to surface under direct questioning. The framing is that a fact can be present in the parameters yet inaccessible to a single forward pass, and that intermediate reasoning steps act as a retrieval mechanism over the model's own memory, raising recall on knowledge the model demonstrably learned during pretraining. It reframes some of reasoning's benefit as internal retrieval rather than fresh computation, which has practical implications for when reasoning helps and when a model is simply confabulating around a genuine knowledge gap.

RoboAtlas is a contextual active-SLAM framework that adaptively trades off geometric exploration against semantic reasoning, built on a scalable 3D semantic mapping system the authors call OpenRoboVox. It combines frontier-based exploration, global reasoning over a semantic map, and egocentric vision-language-model reasoning, so the robot decides where to look next using both where the map is geometrically incomplete and what the scene means. Coupling classical active SLAM with VLM-driven semantic priors is the contribution, aimed at exploration that is efficient about covering space and about finding the semantically relevant parts of an environment.

As spoken conversational systems become common, their ability to read paralinguistic social cues — tone, hesitation, affect — is a bottleneck for natural interaction, yet existing emotional-intelligence evaluations assess reasoning through text alone and miss the acoustic channel. SpeechEQ benchmarks an emotional-intelligence quotient for socially-aware voice models directly on speech, scoring how well systems perceive and respond to emotion conveyed in the audio rather than in a transcript. Measuring emotional competence on the spoken signal itself is the contribution, closing a gap that text-only emotion benchmarks structurally cannot address.

On Latent Space, Databricks co-founders Matei Zaharia and Reynold Xin make the case that the frontier AI ecosystem needs to remain open, arguing from the data-and-infrastructure vantage point that enterprises building on AI are better served by open weights, open tooling, and portability than by lock-in to a single closed provider. The conversation is a useful counterweight to the week's consolidation news — even as labs verticalize around proprietary silicon and guard model IP, a large constituency of infrastructure vendors and enterprise buyers has the opposite incentive, and Databricks is positioning itself squarely on the open side of that divide.

Retrieval-augmented generation is vulnerable to corpus poisoning, where malicious documents inserted into the retrieval index steer model outputs. Existing detection typically bolts on an auxiliary classifier or an extra LLM verification pass, adding compute and latency. This work traces a poisoned target answer back through token-influence attribution — identifying which retrieved tokens disproportionately drove the output — without a separate detector. Attribution-based detection that reuses the generator's own signals, rather than standing up a second model, is the contribution, and it is a more deployable defense for production RAG systems where every extra inference pass costs.

This work constructs what the authors describe as the first two-dimensional hyperbolic neural quantum state, a Lorentz 2D-RNN, and benchmarks it against the Euclidean 2D-RNN on the paradigmatic N-by-N transverse-field Ising model. The motivation is that hyperbolic geometry can represent hierarchical correlation structure more efficiently than flat Euclidean embeddings, and the question is whether that helps a recurrent network approximate many-body ground states. It is a niche but clean test of geometry-aware recurrent architectures applied to quantum many-body simulation, an area where representational inductive bias translates directly into sample efficiency.

Optimization-oriented LLMs that can translate a natural-language problem into a formal optimization model and solve it usually depend on large supervised datasets and costly reasoning traces. MiniOpt targets strong generalization across diverse optimization problem types while using only limited training resources, aiming to reach broad coverage without the heavy data and compute budgets prior approaches assumed. The data-efficiency angle is the point — making optimization-modeling capability reachable for teams that cannot afford to build large bespoke supervised corpora.

As RAG variants like GraphRAG and Agentic RAG proliferate, the practical question is when each is worth its added complexity. This paper offers an evaluation framework that compares regular RAG, GraphRAG, Modular RAG, and Agentic RAG on semi-structured knowledge bases under context optimization, so practitioners can see where the heavier graph- and agent-based approaches actually pay off and where plain RAG suffices. The value is decision guidance rather than a new method — a sober comparison that pushes back on the assumption that more elaborate retrieval machinery is always better.

European governments and industry are pushing back on the United States' approach to the semiconductor export-control fight, signaling friction over how far Washington's restrictions should extend and how they affect European firms and supply chains. The report captures a widening gap between US and European positions on chip-trade policy, which matters for AI because export controls on advanced semiconductors and manufacturing equipment are now a primary lever shaping where frontier compute can be built and sold. Reported here factually for its bearing on the global compute supply chain.