Wolf Digest — Friday, May 29, 2026

Glean's enterprise AI search business crossed $300 million annual run-rate, tripling its top line in the past year even as the hyperscalers (Google with Vertex AI Search, Microsoft with Copilot Search, Amazon with Q Business) shipped competing offerings. TechCrunch's framing: Glean's go-to-market pitch has shifted from find your enterprise knowledge faster to consolidate the seven AI subscriptions your knowledge workers are accumulating. The cross-source signal from Latent Space puts Glean alongside OpenRouter as one of the breakout enterprise-side AI companies of the past two quarters.

A unified embodied foundation model extending Qwen's vision-language stack to continuous action and trajectory generation via a DiT-based action decoder. Trained jointly across robotics-manipulation trajectories, human egocentric demos, simulation data, vision-and-language navigation, and trajectory-centric supervision; embodiment-aware prompt conditioning lets the same model address multiple robot platforms. Manipulation, navigation, and trajectory prediction are unified into one action-and-trajectory output head. Among multi-source HF Daily Papers this run, this is the highest-coverage VLA submission.

The SGLang and AMD teams published a joint post showing competitive total-cost-of-ownership for large-scale DeepSeek-R1 disaggregated inference on Instinct MI355X GPUs, mediated by a new routing-aware kernel orchestrator called MoRI built on SGLang's serving stack. The takeaway is structural: AMD now has end-to-end MoE serving with one of the two leading open-source inference frameworks, closing one of the last remaining ecosystem gaps versus NVIDIA on production inference. Pair with the LMSYS Blog's ROCm-on-Miles post earlier in the month for the RL-training side.

Casts long-horizon LLM interaction as Contextual Belief Management (CBM): maintaining an internal belief state aligned with formal evidence while isolating task-irrelevant noise. Introduces BeliefTrack - Rule Discovery + Circuit Diagnosis with symbolic verifiers for turn-level evaluation. Diagnoses three failures (Failed Stay, Failed Update, Failed Isolation); vanilla LLMs fail systematically, prompts help marginally, and RL with belief-state rewards drops failure rates 70.9% on average. Representation-level steering recovers 46.1% in the residual.

Generative multi-agent world model for interactive simulation beyond two players. Introduces Simplex Rotary Agent Encoding - a parameter-free extension of 3D RoPE placing agents as simplex vertices in rotary angle space, giving distinct phases while keeping permutation equivalence and supporting scalable identity without per-slot learned embeddings. Sparse Hub Attention reduces cross-agent attention cost via learnable hub tokens. The named target is interactive video generation that needs to scale to many simultaneous controllable agents.

NVIDIA's ICRA roundup covers its 2026 robotics research portfolio with the explicit framing of moving from controlled-demo and scripted-automation regimes toward generalizable real-world embodied autonomy. Key threads: sim-to-real transfer with massively-parallel GPU-accelerated rollout; foundation models for robot policies that can be specialized with little data; and Isaac Lab tooling that absorbs the sim-to-real engineering rather than leaving it to research labs. Useful as the company-line version of where NVIDIA wants the robotics stack to land - pairs cleanly with the Physical Intelligence and Boston Dynamics announcements from earlier in the month.

The Pentagon acknowledged that US service members have been targeted using commercially-available location data sold by data brokers. The C4ISRNET piece details an operational-security pattern in which adversaries purchase precision-located mobile-ad data and cross-reference with installation perimeters. While not directly an AI story, the targeting pipeline relies heavily on automated cross-referencing and is the cleanest data-broker-as-weapon example surfaced publicly to date - relevant for any policy-side conversation about commercial AI-derived data flows.

Latent Space's interview with Cognition's Walden Yan and OpenInspect's Cole Murray maps the practitioner architecture of async background agents: how teams structure background-agent harnesses, where Anthropic's managed-agents and Google's Gemini managed agents fit relative to LangGraph and Pydantic-based DIY stacks, and the operational realities of running agents at Shopify, Stripe, and elsewhere. Yan's framing in the conversation is that the central tension in agent-product strategy is between decacorn agent labs (Sierra, Decagon, Notion, Cursor) and the increasing ease of DIY - and async background agents are where DIY currently wins.

Identifies a Thinking-Acting Gap in agentic RL: vanilla GRPO sees tool calls attempted in only ~30% of rollouts and all-wrong in ~40% of attempts, gutting the learning signal at exactly the steps that need it. AXPO fixes the thinking prefix and resamples the tool call + continuation for each all-wrong tool-using subgroup, with uncertainty-based prefix selection. Across 9 multimodal benchmarks x 3 scales of Qwen3-VL-Thinking, SFT+AXPO beats SFT+GRPO by +1.8pp Pass@1 / Pass@4 at 8B on average; 8B+SFT+AXPO surpasses the 32B Base on Pass@4 at 4x fewer parameters.

Lightweight agent-safety alignment framework targeting open-world agents like OpenClaw. Updates agent-safety taxonomy for Codex/OpenClaw execution scenarios, builds a taxonomy-guided data engine with influence-function purification, trains AgentDoG 1.5 variants at 0.8B/2B/4B/8B parameters using only ~1k samples and gets comparable performance to closed-source baselines (e.g., GPT-5.4). Deployable as a Docker-level guardrail at SOTA latency with two orders of magnitude lower overhead than the alternatives.

TechCrunch's long-read frames the shift from human-driven to agent-driven traffic as a structural rebuild rather than a routing tweak: AWS, Cloudflare, and others are explicitly redesigning the cloud edge for traffic that originates from agents - different cache patterns, different rate limits, different attestation needs (machine-to-machine identity, agent-readable robots.txt extensions, AI-traffic carve-outs in DDoS heuristics). Noteworthy data point: machine-generated traffic crossed half of total web traffic in 2025 per Cloudflare's Radar; the projection is 70% by end of 2026.

OpenAI's case study with the IT services firm Endava reports that Codex compressed requirements analysis from weeks to hours and is now used across the firm's delivery pipeline. The publication is squarely a Codex-marketing artifact, but the operational detail (specific stages of the Endava delivery model that Codex now owns, automation of UML/architecture-diagram generation upstream of code synthesis) is the kind of system-design signal worth tracking - IT services firms with code-as-product remain a leading indicator of enterprise codegen adoption.

Proactive Recommendation via Rectified Policy Gradient. Identifies two flaws in naive RL for proactive recommendation: positive-mean step-decomposition creating length-dependent gradient bias, and high-variance path-level reward weighting. Stepwise Reward Centering neutralizes length bias; Position-Specific Advantage Estimation uses decomposition structure to compute step-dependent advantages. The methods package targets RecSys problems where path-length blowup has historically broken policy optimization.

Native vision-language model that learns cross-frame and pixel-word correspondence end-to-end without external image encoders or post-hoc fusion. By eliminating module boundaries, NEO-ov enables fine-grained spatiotemporal modeling to emerge inside the model. Closes the gap to modular VLMs while excelling at fine-grained visual perception, with detailed training recipes published. Architectural argument: native one-vision designs are competitive at scale.

14,056 research-level math problems curated via multi-agent pipeline - largest such dataset to date - plus 220K teacher trajectories from two open models. Reports a striking failure mode: newer model generations produce 5.6x more references and 5.0x more *fake* references per trace. After agentic filtering, fine-tuning Qwen3 from 4B to 30B yields +9.2 points average gain over base. Argues filtered open-problem attempts give useful supervision even without correct reasoning traces.

Error tracing and attribution for LLM memory systems. Transforms memory pipelines into executable memory-evolution graphs enabling fine-grained tracing of operational information flow. Companion benchmark MemTraceBench covers Long-Context, RAG, Mem0, EverMemOS; automatic attribution iteratively traces operation subgraphs to root-cause failed cases. Demonstrates that memory failures are systematic at the operation level (information loss, retrieval misalignment) and uses the attribution signal to guide prompt optimization in a closed-loop self-correcting system.

Asana announced an acquisition of StackAI, a no-code agent-builder, to expand its in-app AI workflow tooling. Terms undisclosed. The acquisition is part of a broader trend - work-management platforms (Asana, Monday, ClickUp, Notion) competing to embed agentic builders rather than ship MCP integrations to third-party agent platforms. Notable mainly as a data point on consolidation among second-tier work-management vendors trying to keep parity with Notion's agentic surface area.

Google's I/O 2026 research recap consolidates the platform-side announcements - Gemini 3.5 Flash specifically - and the research artifacts that landed across the conference. Notable for context: Gemini 3.5 Flash now sits at 55.3 on the Artificial Analysis Intelligence Index and 210 tokens/s, putting it firmly in the high-end-speed quadrant. The Google AI blog post itself is recap-format and light on technical detail; treat as a pointer to the underlying I/O sessions.

Defense One reports the Navy used uncrewed surface vessels and aerial drones to sink a retired warship during a live-fire test, validating an end-to-end maritime kill chain with autonomous platforms. Pair with the prior month's announcements from the Pentagon's Replicator initiative and the Manned-Unmanned Teaming Pentagon framing - the cadence of autonomous-platform live-fire tests has accelerated.

Quantifies LoRA's exact parametric memory capacity as a controlled probe in latent space. Introduces the Parametric Memory Law - a power-law linking loss reduction (delta-L) to effective parameters and sequence length. Token-level analysis reveals a deterministic phase transition: p>0.5 is sufficient for verbatim recall under greedy decoding. MemFT redistributes training budget toward sub-threshold tokens, improving memory fidelity and efficiency.

Local Modality Substitution for deeper vision-language fusion. Diagnoses 'carrier sensitivity' in VLMs - replacing a textual question with a rendered-image counterpart should be neutral but in practice causes large performance drops, tracing back to asymmetric text-as-query / image-as-reference roles in training data. LoMo is a lightweight architecture-agnostic data curation pass that supplies supervision for representation alignment across textual and visual carriers.

GASP injects geometric priors into VLM transformer layers. Small correspondence head applied as deep supervision across all layers; dual objective: contrastive loss on ground-truth point correspondences (2D view-invariance) and depth-consistency supervision (3D ambiguity resolution). Reports that standard VLMs have layer-wise correspondence-matching accuracy below 5%; GASP substantially improves peak layer-wise correspondence and downstream 3D reasoning.

Text-to-FHIR pipeline producing clinically-realistic HL7 FHIR R4 bundles from unstructured text via staged LLM generation + terminology-grounded validation/repair. Achieves valid FHIR for 82.5% of MedCaseReasoning cases. Headline finding: LLMs score consistently lower diagnostic accuracy on structured FHIR inputs than on the plain-text equivalents, an EHR-aligned benchmark caveat that should reshape clinical-AI evaluation practice.

Entropy-Cut Metropolis-Hastings - a sampler for the power distribution of base LMs that targets decision points instead of uniform cuts. Uses base-model next-token entropy as a proxy for key decision points and resamples from there, fixing the failure mode where uniformly-chosen cuts rewrite local details rather than revisit consequential choices. Provides RL-free reasoning at parity with trained reasoning models.

Substitutes external supervision with recovery-oriented optimization from weak-model failures. Trains directly on incorrect reasoning traces as opportunities for improvement; outperforms strong on-policy RL baselines on competitive math and general reasoning, with stronger self-corrective behavior as training difficulty rises. Architectural appeal: scalable RL without distilled teacher models or curated difficulty datasets.

Causality lens on video generation as world modeling. Distinguishes generative pattern-matching from causal scene modeling and proposes a counterfactual evaluation protocol. Most current video generators succeed at temporal-pattern matching while failing causal counterfactual tests at high rates; the gap quantifies how far video generation is from being a true world model rather than a sequence prior.

Characterizes backdoor attacks injected via LoRA adapters: token-level trigger generalization, attack-surface mapping across adapter ranks, and detection heuristics for hostile LoRA artifacts in adapter-shared marketplaces. Worth flagging given the increasing reliance on community LoRA distribution in the Hugging Face ecosystem.

Do LLM agents exhibit human-like psychology? HEART-Bench probes agents across emotional, social, and cognitive dimensions, finding systematic deviations from human baseline psychology while also surfacing instrumental cases where human-like response patterns emerge from instruction-following rather than internalized models. Useful as a methodological corrective to studies that conflate output-similarity with cognitive parallelism.

Bidirectional Evolutionary Search frames self-improvement as joint forward generation + reverse refinement passes. Demonstrates exponential reductions in samples required to escape the entropy shell that bounds best-of-N improvement, with empirical results across reasoning benchmarks at scale.

Autonomous-research agent built on Chain-of-Evidence - iteratively decomposes scientific questions into evidence-yielding sub-investigations, grounds claims to literature/data sources, and composes findings. Targets human-level autonomous research as a north-star and reports concrete progress on multi-paper synthesis tasks.

Empirical study: AI research agents systematically narrow the scientific search space relative to human researchers, biasing exploration toward already-cited and well-resourced subfields. Quantifies the narrowing across multiple domains and proposes diversification heuristics for agent deployment in scientific workflows.

Optimizes agent skills via a gradient-like update rule over textual skill definitions. Skills are treated as the natural-language analog of parameters; updates flow from skill-execution evals back to skill text via local perturbation + acceptance. Reports gains over prompt-only baselines on agentic benchmark suites without touching model weights.

Unified retrieval across heterogeneous knowledge sources - code, structured data, unstructured documents, multimodal - in a single retrieval substrate. Architectural argument that source-aware encoder zoos can be collapsed into a shared latent without losing source-specific recall.

Replaces discrete memory-slot abstractions with a continuously-evolving connectivity graph where retrieval and update are local message-passing operations over a sparse neighborhood. Reports better performance on long-horizon memory benchmarks than slot-based predecessors.

Coins 'agent lifespan engineering' for the operational problem of deployed agents that accumulate state, drift behaviorally, and require curated retirement / refresh cycles. Provides an empirical study of state-accumulation patterns in long-running agents and design patterns for graceful state pruning.

Generative Supervision Helps Embodied Intelligence - uses generative models of action trajectories as auxiliary supervision for embodied policies. Reports gains on standard manipulation benchmarks at lower sample count than direct behavior-cloning, with the generative trajectory model serving as both data augmentation and consistency regularizer.

Automated domain specialization for small computer-use agents. Identifies a small-agent's failure modes via comparison to a stronger reference, then synthesizes targeted training data to plug each weakness. Yields domain-specialized small agents that outperform untargeted fine-tuning at substantially lower data and compute cost.

Dynamic, multilingual, multi-domain benchmark for misinformation detection. Crucially refreshes its corpus from community fact-check feeds rather than fixing a static set, attacking the staleness problem that plagues prior misinformation benchmarks.

Preplan-Empowered LLM Mathematical Reasoning separates the what-to-solve planning step from the how-to-solve execution step, training a model to produce an explicit plan before tackling math problems. Reports consistent gains on contest-level benchmarks against direct-reasoning baselines.

Recoverable Program-of-Thought with checkpoint repair. Adds checkpointed intermediate state to Program-of-Thought reasoning so the model can recover from a fault without restarting; improves robustness on multi-step computational reasoning benchmarks.

Major exchanges are designing derivative products around AI inference tokens, treating them as a raw input commodity comparable to electricity or bandwidth. The structural argument is that token pricing has become dense and liquid enough across providers (with cache-hit vs. input vs. output pricing already standardized) to support futures and options contracts hedging enterprise inference spend. Noteworthy as a leading indicator of inference-economy financialization rather than as a near-term product.

Cohere published three May-28 enterprise-facing posts (AI in BI use cases; a practical guide to Model Context Protocol; AI governance for safe enterprise adoption) alongside a May-27 Mila partnership for Quebec French. Most interesting in the context of last week's Command A+ open-source enterprise model launch: the MCP post is a tacit acknowledgement that MCP has won the standard-for-agent-tool-calling slot, and Cohere wants to be visibly aligned with it.

The Army Data Operations Center, stood up in April, is being staffed with minimal human headcount on the explicit expectation that automation will pick up growing force-wide data-ops demand. DefenseScoop's piece pairs the staffing plan with the standing-up of Army-side AI data pipelines - a clean case of an institution betting that LLM-driven automation will close a workforce gap rather than waiting for evidence.

Are search agents searching, or just verifying what they already believe? LiveBrowseComp is a benchmark designed to detect the failure mode in which a browser agent's retrieval queries chase confirmatory evidence rather than expand the hypothesis space. Reports the prevalence of confirmation-style queries across leading search agents.

Canonical-Context On-Policy Distillation tackles inconsistency between teacher and student when given the same evidence but different contexts. Forces both into a canonical context before the distillation step, recovering most of the consistency gap and improving downstream accuracy without re-collecting trajectories.

Early-stops rollout length during on-policy distillation when the student's per-token KL plateaus. Eliminates a substantial fraction of redundant token computation with no measurable downstream loss.

Benchmark for robotic creative problem solving - surfacing the unexpected ways that embodied agents fail at tasks requiring genuine insight rather than pattern-matching. Robotic agents systematically struggle with creative re-purposing of objects, with failure modes that map cleanly onto missing-world-knowledge rather than control-stack limitations.

Evaluates LLM scientific hypothesis generation under progressive context degradation. Headline finding: LLM hypothesis quality degrades non-monotonically as context is degraded, with sharp cliff-edges at certain degradation thresholds rather than a smooth roll-off - argues for evaluation that includes adversarial context degradation.

Latent-reasoning study showing LLM working memory has structured latent representations that can be unlocked with targeted intervention. Reports that residual-stream patches at specific layers expand effective working memory capacity by a measurable fraction without retraining.

Information-Bottleneck Driven Tree-based Policy Optimization. Uses an information-bottleneck regularizer over a tree of policy rollouts to balance exploration and exploitation in long-horizon agentic settings.

Benchmarks thinking-mode switch strategies in hybrid models that mix fast and slow reasoning. Reports the operational tradeoffs of various switching policies - confidence-threshold, query-difficulty heuristics, learned routers - across reasoning workloads.

Uses SAE-derived features as a guide for post-training data engineering - selects training data that activates SAE features the target model needs to strengthen, rather than relying on quality or diversity heuristics over text. A practical use of interpretability artifacts for training-data curation.

Collects 50 effects in 1 LoRA via Multi-Teacher On-Policy Distillation. Demonstrates that a single low-rank adapter can encode many distinct stylistic effects when distilled from a panel of teacher models, dramatically reducing the per-effect LoRA proliferation typical in generative-media tooling.

Human-like long-document translation agent with observe-and-act adaptive policy. Decomposes long-document translation into observe-translate-revise steps with explicit cross-document coherence checking. Reports gains over GPT-4-class single-pass translation on book-length corpora.

Coordination-policy substrate for multi-agent systems. Provides a learnable substrate for inter-agent message passing and coordination policy that separates per-agent reasoning from system-level coordination.

Microsoft Research's Data Formulator 0.7 release adds enterprise-data connectors and AI-driven analytic flow construction on top of the existing visual chart-building pipeline. The release positions Data Formulator as Microsoft's research-backed alternative to the Tableau / Power BI Copilot story for analyst-facing workflows. No frontier-model news here, but a useful signal on the enterprise-tooling layer that wraps Copilot.

Efficient high-quality video generation with sparse sequence parallelism. Targets the throughput bottleneck of large-resolution video diffusion at long durations; reports throughput improvements at parity quality.

DPO recipe for post-trained LLMs that recovers response diversity without sacrificing alignment metrics. Targets the well-documented diversity collapse from preference optimization; introduces a diversity-aware preference-pair construction that preserves alignment scores while improving output diversity by measurable margins.

Long-horizon LLM agent training with meta-cognitive memory policy optimization - explicit RL over both action-policy and memory-update-policy. Reports stronger long-horizon performance than memory-static baselines on multi-day-task benchmarks.

a16z general partner Erin Price-Wright argues policymakers' rebuilding-industrial-base rhetoric requires AI as the enabling layer - that the gap between needing to build and being able to build is bridged by automation, robotics, and AI-driven design. The post is policy-positioning material for an a16z faction increasingly focused on defense-industrial and reshoring policy, and is worth tracking as the venture firm's house-line on AI's policy framing heading into the FY27 NDAA cycle.

MIT Tech Review covers a new direct-lithium-extraction process targeting geothermal brines, with implications for the battery-supply chain that backs both EVs and AI datacenter power buffering. Not strictly AI but tracked here because lithium availability is a load-bearing assumption in the 5GW datacenter expansion announcements that dominated this week's news.