Wolf Digest — Thursday, June 18, 2026

Odyssey, a world-model startup founded by self-driving veterans Oliver Cameron (ex-Voyage, Cruise) and Jeff Hawke (ex-Wayve), raised a $310 million Series B at a $1.45 billion valuation led by Natural Capital, with Amazon, AMD Ventures, and GV participating. Odyssey builds world models that gather physical-world data — it has people walk around with backpack-mounted cameras, Google Street View style — and simulate it with accurate physics, and is best known for generating rich, interactive video from text prompts. As part of the round, AWS becomes Odyssey's preferred cloud and the startup will optimize its models for AWS Trainium chips, a competitor to Nvidia's accelerators.

LoopCoder-v2 is a family of 7B parallel-loop Transformer (PLT) coders trained from scratch on 18T tokens, studying loop-count selection under a gain–cost view: extra loops refine representations but cross-loop position offsets introduce a positional mismatch. The two-loop variant broadly beats the non-looped baseline across code generation, code reasoning, and agentic software engineering, improving SWE-bench Verified from 43.0 to 64.4 and Multi-SWE from 14.0 to 31.0. PLT uses cross-loop position offsets and shared-KV gated sliding-window attention so loop count becomes a practical test-time-compute knob.

OmniAgent reframes long-video understanding as active perception rather than the uniform “watch-it-all” paradigm, whose cost grows with video duration. It is presented as the first native omni-modal agent that decides what to attend to instead of relying on global pre-scanning, keeping context cost from scaling with video length and improving accuracy on long-video benchmarks.

The Marine Corps activated Marine Unmanned Maintenance Squadron 14 (MUMS-14) at MCAS Cherry Point, its first organic unit dedicated to sustaining large drones, specifically the MQ-9A Reaper fleet. The roughly 300-person squadron — mostly UAV technicians, mechanics, and ground-control-station maintainers — replaces the contracted logistics support that forward-deployed Reaper detachments had relied on. The activation lands alongside a bipartisan right-to-repair provision, championed by Senators Warren and Sheehy, included in the Senate Armed Services Committee's 2027 NDAA markup that would let troops fix their own equipment without contractor lock-in.

RLVR methods like GRPO commonly suffer policy-entropy collapse during training. A first-order gradient analysis of token-level entropy dynamics identifies a token-level credit-assignment mismatch; STARE reweights advantages at the token level using surprisal to stabilize entropy, preserving exploration without the usual collapse during reasoning-model post-training.

Post-training of reasoning models leans on distillation from chain-of-thought annotations that are costly and often noisy, and on RLVR with sparse signals. This work conditions self-distillation on rubrics to provide denser, more reliable supervision than imperfect rationales, aiming to combine distillation's stability with reward-aligned correctness.

GitHub Copilot's Auto mode now routes requests through a custom model that chooses among available LLMs based on reasoning depth, code complexity, debugging difficulty, and tool-orchestration needs, per a GitHub blog post and an accompanying research paper. The pitch is getting more out of each token by matching task to model rather than defaulting to one, part of a broader move toward learned model-routing inside coding agents.

Preference-based RL learns reward models from pairwise behavior comparisons but typically uses passive data collection with poor early sample efficiency. UBP2 is a model-based approach that actively directs exploration by jointly reasoning over uncertainty in the reward model and the dynamics, improving sample efficiency on control and robotics tasks.

OneCanvas gives VLMs spatial reasoning without model-specific geometry encoders or large training budgets. It unprojects each patch to 3D world coordinates using depth and camera pose, then aggregates all views onto a single equirectangular panoramic canvas, a lightweight representation for 3D scene understanding usable in embodied settings.

TherapeuticsBench Preclinical Pharmacology (TxBench-PP) is a verifiable benchmark for small-molecule preclinical pharmacology, the first slice of a broader TherapeuticsBench effort. It targets realistic drug-program decisions to test whether AI agents can be trusted to compress interpretation-and-decision loops in drug discovery, with grounded, checkable evaluation.

Uniform diffusion language models (UDLMs) let any token update at any step, in principle enabling more flexible generation than masked diffusion or autoregressive decoding, but none had been pretrained from scratch at both large parameter and token scale. Sumi is an open UDLM trained at scale, providing a comparison point for diffusion-based LM generation against established autoregressive and masked-diffusion models.

Deploying multimodal foundation models as closed-loop policies increasingly requires conditioning actions on observations that are no longer visible. RNG-Bench (Reconstructive Non-Markov Games) isolates a base model's ability to maintain and act on hidden state, separating it from full-state exposure and post-episode recall that existing benchmarks conflate.

Standard RLVR samples responses independently and optimizes from final answers, causing redundant exploration of similar intermediate reasoning and wasting compute under sparse rewards. GraphPO structures responses as a graph to share intermediate reasoning and densify credit assignment, improving the efficiency of policy optimization for large reasoning models.

Long-context reasoning is essential for agentic LLMs reasoning over lengthy trajectories. Arguing that prior RL work overfocuses on reward engineering while diverse training data stays scarce, this paper revisits long-context RL from a data-centric angle, proposing a recipe for assembling training data that improves long-horizon reasoning.

On-policy self-distillation trains a model on its own rollouts with a frozen copy giving dense token-level targets, but extending it to multimodal models opens a shortcut: the privileged target can guide tokens from the text reference rather than the image. ViGOS forces visual grounding so the dense targets depend on the image, decoupling perception from text-driven reasoning shortcuts.

LLMs reason well over symbols but are blind to quantum representations such as unitary matrices. This work maps unitary operators into an LLM's latent space, enabling unified modeling over quantum and linguistic inputs and a step toward letting language models reason about quantum operators directly.

Embodied VLA models are made by fine-tuning pretrained VLMs on robotics data, but it is unclear how much commonsense and factual knowledge survives. Failures on knowledge-sensitive tasks conflate missing knowledge with poor low-level control. Act2Answer is a lightweight protocol adapting VLM knowledge benchmarks to VLAs, isolating retained knowledge from control generalization.

A study of sparse-autoencoder feature interventions finds they are unreliable: after a feature is suppressed, downstream computation partially recovers the suppressed information, undermining causal claims about SAE features as clean control knobs. The result is a caution for interpretability work that treats SAE interventions as faithful, surgical edits to model behavior.

Inference platform Fal rolled out Kling 3.0 Turbo and related Omni upgrades: faster generation, lower cost, better lip-sync, more stable motion, stronger prompt and reference consistency, clips up to 15 seconds, and full 4K generation with Omni, plus improved storyboard and multishot workflows. The update is an incremental but practical push on the cost-quality-speed frontier for text- and image-to-video generation.

Fixed spike encoders force downstream spiking neural networks to compensate for non-adaptive inputs, a bottleneck for neuromorphic speech. This work presents a learnable residual speech-to-spike encoder trained end-to-end with a recurrent SNN backbone, adapting the encoding to the acoustic signal and improving event-driven speech processing.

Simulating human users could advance agent-assistant training, personalization evaluation, and social-science research. Rather than matching a single ground-truth response by log-probability or similarity, Turing-RL uses a Turing-test-based reward so the simulator is trained to be indistinguishable from real users, capturing the diversity of plausible human behavior.

Score- and flow-matching generators often lean on preference-based RL both to align with subjective preferences and, oddly, to recover realism and coherent object structure that matching training should already learn. Arguing this reflects a structural mismatch, the paper corrects flow matching with discriminators to recover those properties directly from data rather than via preference optimization.

Hybrid architectures interleave softmax attention with cheaper linear or recurrent mixers to cut quadratic cost. This work re-examines what the efficient-attention component actually contributes in such hybrids, clarifying where linear-attention layers help and where full attention remains necessary for quality at long context.

Industrial 10B-level foundation models set the bar for image inpainting but are too costly to deploy. Moebius is a 0.2B lightweight inpainting framework that targets 10B-level performance by addressing the representation bottleneck that extreme structural compression usually triggers, offering a deployable specialist alternative to giant generalists.

Cohere published an engineering deep-dive on fair scheduling for multi-tenant LLM inference, aimed at preventing a single heavy tenant from monopolizing shared GPU compute and starving others — the classic “noisy neighbour” problem. The post describes mechanisms to give each tenant its fair share of serving capacity under contention, a practical concern as inference platforms pack many workloads onto shared accelerators.

The SGLang-JAX team added efficient serving of inclusionAI's Ling-2.6-1T (a trillion-parameter MoE) on TPU v7x. Profiling pinned the mixture-of-experts path as the main bottleneck; the team hides MoE data movement behind compute using a single Pallas kernel, a concrete TPU-serving optimization for very large sparse models.

Most agent memory benchmarks assume a single user, leaving shared assistants for hospitals, workplaces, and households understudied. In those settings many principals write to a common memory pool and query it under different roles and scopes, so memory quality requires governance, not just recall. GateMem benchmarks multi-principal shared-memory agents on access control and role-scoped retrieval.

Anthropic announced a Seoul office and a set of partnerships across the Korean AI ecosystem, extending its Asia-Pacific enterprise and public-sector footprint. The move continues Anthropic's international expansion and regional go-to-market buildout following recent partner-network and enterprise-integration announcements.

LMSYS, with MOSI and the OpenMOSS team, announced end-to-end serving for MOSS-TTS-Local-Transformer-v1.5 on SGLang-Omni. The open text-to-speech model produces native-streaming 48 kHz audio, targeting low-latency conversational voice agents on an open serving stack.