Wolf Digest — Monday, June 15, 2026

MRAgent replaces the static retrieve-then-reason paradigm with an associative memory graph plus active reconstruction. Memory is stored as a Cue-Tag-Content graph in which associative tags bridge fine-grained cues to content, and the agent interleaves LLM reasoning directly into memory access, iteratively reconstructing relevant context from intermediate evidence discovered during inference rather than fetching a fixed top-k up front. The framing — reconstruction over retrieval — targets the long-horizon reasoning failures of memory-augmented agents.

This work pushes retrieval-augmented generation into long egocentric video, where systems must select query-relevant chunks across modalities and temporal granularities. It introduces V-RAGBench, a benchmark of query / evidence-chunk / answer triplets that decouples retrieval from generation (existing benchmarks let queries be answered without the video, hiding retrieval errors), and CARVE, which runs parallel retrievers across configurations with chunk-adaptive reranking rather than one fixed modality-granularity setting per query.

A group of state attorneys general has opened an investigation into OpenAI, according to TechCrunch. The specific states involved were not disclosed, but the inquiry reportedly spans a wide range of issues, from the company’s advertising policies to its handling of health data. The probe adds to mounting regulatory scrutiny of frontier labs and signals that state-level enforcement, not only federal action, is becoming a live channel for AI oversight in the United States.

Orchestra-o1 is an omnimodal agent-orchestration framework for tasks that require unified understanding and coordination across text, image, audio, and video. The authors argue existing orchestration frameworks handle only a narrow set of modalities and struggle when heterogeneous modalities coexist and interact; Orchestra-o1 targets efficient task decomposition and collaboration in these omnimodal settings, extending multi-agent orchestration beyond the text-centric regime.

OpenAI introduced a Partner Network backed by a 150-million-dollar investment to help consulting and systems-integration partners accelerate enterprise AI adoption, deployment, and transformation. The move is a distribution play rather than a research one: it formalizes a channel for enterprises that need implementation support around GPT-5.5-class models, and mirrors the partner-ecosystem strategies that cloud vendors have long used to scale adoption. The timing, amid a tightening policy environment around frontier-model access, positions OpenAI to capture enterprise demand through services partners.

OmniDirector clones camera motion from reference videos for controllable video generation. It encodes cameras as grid-motion videos — a visual representation of camera parameters that supports integrating diverse trajectories for multi-shot generation — and trains on a million-scale camera-grid/video corpus, avoiding the parametric representations that fail on multi-shot settings and the scarce, synthetic cross-paired data prior methods relied on. The result is more robust cloning of complicated, multi-shot camera motion.

HarnessX treats the agent harness — prompts, tools, memory, and control flow — as a first-class object to be composed and improved rather than hand-crafted per model and task. It assembles typed harness primitives via a substitution algebra and adapts them through AEGIS, a trace-driven multi-agent evolution engine that mirrors symbolic adaptation and reinforcement learning, distilling execution traces back into systematic harness improvement and closing the harness-model loop.

A Two Minute Papers walkthrough highlighted NVIDIA’s Nemotron 3 Ultra, an open-weight model (roughly 550 billion total parameters with about 55 billion active) that on Artificial Analysis’s independent benchmarks combines a competitive Intelligence Index score with class-leading output speed near 183 tokens per second. The release continues NVIDIA’s push to pair open weights with strong inference efficiency, positioning Nemotron as an openly available alternative for builders who want frontier-adjacent capability without API access constraints.

This report investigates how machine intelligence might develop beyond human-level AGI toward artificial superintelligence, using the theoretically well-understood endpoint of Universal AI to provide formal grounding for analyzing the AGI-to-ASI transition. It is a conceptual and forecasting contribution rather than an empirical result, mapping the questions a post-AGI trajectory raises; its value lies in framing the continuum rigorously, and its limits in the speculative nature of any claims about systems that do not yet exist.

This paper identifies policy-level (rather than token-level) diversity as a lever for GRPO. Smaller models in the same family show higher policy-level diversity — superior pass@k as sample counts grow — that is temporally correlated and preserves logical consistency, unlike the step-wise noise injected by token-level randomness. The proposed S2L-PO (Small-to-Large Policy Optimization) uses smaller models as structured explorers to supply cleaner exploration signal for gradient estimation in larger ones.

Simon Willison flagged the Pyodide 314.0 release, which lets package maintainers build and publish Pyodide-compatible WebAssembly wheels directly to PyPI under the PEP 783 PyEmscripten ABI, installable at runtime in the browser. Previously the Pyodide maintainers had to build and host more than 300 packages themselves, a major community bottleneck. The change matters for browser-side and agent tooling: it makes the long tail of Python packages installable in client-side runtimes without a separate hosting pipeline.

Most coding agents consume repositories as plain text, unlike human developers who exploit visual structure such as folder hierarchies and dependency graphs. This first systematic study evaluates four multimodal models on repository-level issue resolution using visual repository representations. A key finding: a strictly vision-only setup degrades accuracy, implying visual structure helps as a complement to, not a replacement for, textual repository context.

OmniVideo-100K is a dataset built to fix the decoupled 'video-caption-QA' pipeline, which segments video into short clips and describes audio and visual streams separately — severing sound-to-source associations and producing inconsistent entity descriptions across segments. It uses structured scripts and explicit evidence chains to keep audio-visual associations intact for long-form audio-visual question answering, supporting reasoning that spans both modalities coherently.

Applying GRPO to GUI grounding suffers a degenerate-group problem: rollouts from a single screenshot view often come out all-fail on hard instances or all-success on easy ones, yielding no useful relative advantage. VISTA builds each comparison group from multiple target-preserving views of the same GUI instance — crops that keep the target element visible — restoring informative within-group variance and giving a self-verified training signal for grounding.

HyVLA-0.5 is an end-to-end report spanning the full robot-learning stack: data collection, model design, continued pre-training and supervised fine-tuning, reinforcement-learning post-training, and real-world deployment. It is presented as an integrated system in which each component plays a distinct role, offering a reference recipe for taking a vision-language-action model from data to a deployed real-world robot policy.

This work shows that pretrained transformer layers can be treated as modules and dynamically skipped or looped per input — a training-free 'program-of-layers' (PoLar). For many inputs, substantially shorter layer programs match or beat full-depth execution, and some originally-incorrect predictions are corrected by alternative programs. The finding points to input-adaptive depth as a free efficiency and accuracy lever in existing models, no retraining required.

A TechCrunch piece surveys the widening rush of AI companies toward public markets and the suppliers, infrastructure providers, and smaller startups trying to ride the wave. The framing — startups hoping to catch a SpaceX-style IPO surge — captures a market in which capital is increasingly available for AI exposure, even as the same week’s export-control news underscores how policy risk now sits alongside the financial upside.

Agent execution traces expose rich procedural detail — tool invocations, intermediate decisions, error-recovery logic — that can leak private skills such as formulas, thresholds, and strategies without any access to model weights. RedAct introduces CapTraceBench (75 specialized long-horizon tasks) to quantify this trace-based skill-extraction risk and evaluates redaction methods that preserve observability and accountability while limiting what downstream parties can reconstruct.

Treating LoRA’s scaling factor alpha as a mere complement to the learning rate misses its real role. Through empirical analysis and a Signal-Drift theoretical framework, the authors show alpha is the dominant driver of effective optimization, delivering gains that learning-rate scaling alone cannot replicate. The practical upshot is that tuning alpha deserves first-class attention when fitting low-rank adapters, not treatment as a fixed hyperparameter.

Most medical-hallucination benchmarks focus on data collection but ignore where errors originate. ClinHallu enables source-level diagnosis, distinguishing hallucinations that arise from visual misrecognition, incorrect medical-knowledge recall, or flawed reasoning integration. By localizing the failure stage within a multimodal model’s reasoning, it gives a more actionable picture for building trustworthy clinical decision-support systems than aggregate accuracy alone.

High medical-licensing-exam scores do not imply safe judgment. This paper shows that when misleading context is injected into questions an LLM originally answers correctly, it frequently abandons the correct answer. The authors term the ability to hold correct judgment under adversarial context epistemic resilience and introduce MedMisBench to measure it — a pointed caution as patients increasingly rely on chatbots for health advice.

CacheRL trains small agent foundation models that reach 92 percent process accuracy on multi-step tool-calling tasks — approaching GPT-5’s reported 94 percent — while using about 100 times less compute. It tackles three practical problems: transferring tool-calling knowledge from large models at scale, running reinforcement learning without costly live tool execution, and learning robustly from noisy cached environments, via cached rollouts and a hybrid reward.

KPMG withdrew a published report on AI usage after apparent fabrications were found in the document, TechCrunch reported. The episode is another instance of an AI-assisted professional deliverable being retracted for hallucinated content, reinforcing that even sophisticated enterprise users continue to ship unverified model output. For practitioners it is a reminder that retrieval grounding and human review remain load-bearing in high-stakes reporting.

In typical text-to-image systems LLMs only encode text while a separately trained backbone does denoising. RepFusion leverages representation autoencoders, which shift the generation target toward semantically structured visual representations more compatible with pretrained LLM priors, and — inspired by how a simple MLP projector aligns visual representations in multimodal LLMs — brings those priors into the denoising step itself rather than only the text encoder.

Video world models can synthesize high-fidelity sequences but often fail to maintain a stable internal state over long horizons. MBench argues existing benchmarks over-weight visual quality, motion coherence, and text-video alignment while overlooking memory — the core requirement of a functional world model — and introduces a comprehensive suite to measure whether a model preserves and reasons over its state across extended temporal spans.

Diffusion-transformer video models are bottlenecked by the quadratic cost of 3D attention. Rather than only cutting per-step compute via sparse attention or KV-caching, RhymeFlow relaxes the rigid constraint that every frame advances through the same denoising step in lockstep, scheduling denoising asynchronously across frames. The training-free approach reduces latency while preserving the standard pipeline’s quality.

SciAgentArena evaluates AI agents in realistic research scenarios, addressing two gaps: agent benchmarks rarely capture the complexity and extended reasoning of scientific work, while scientific-task benchmarks reduce research to static problems with little interactive evaluation. It comprises roughly 200 tasks with stepwise verification inside an interactive, agent-agnostic environment, aiming to measure practical research capability rather than one-shot answer accuracy.

Activation steering is a lightweight inference-time control method, but its success depends heavily on prompt, concept, model, and configuration, usually requiring expensive grid searches and post-hoc rollout evaluation to find the working regime. This work investigates whether steerability can be predicted from a model’s internal states at the very start of generation, which would let practitioners anticipate where steering will and will not take hold without exhaustive search.

Avatar V targets avatars that are not only visually similar but behaviorally recognizable — reproducing a person’s talking rhythm, gestural tendencies, and expression dynamics. The authors argue single static-image conditioning provides insufficient identity and motion information, and that pixel-level objectives underserve the perceptually critical facial regions; Avatar V instead conditions on reference video and scales video-reference avatar generation to capture dynamic motion traits.