Today's papers cluster around turning source code itself into training fuel for coding and computer-use agents, with several groups building large synthetic skill or environment banks directly from GitHub repositories. A second cluster targets efficiency at the architecture level: block-conditioned mixture-of-experts, context-aware memory lookups, and calibrated distillation all aim to decouple capacity from compute cost. On the reliability side, one paper documents a recursive collapse risk when AI-generated peer reviews start training future AI reviewers. Meanwhile the news cycle turned skeptical of agent autonomy: Amazon blocked Meta's Muse shopping agent, a UN panel called for AI safeguards ahead of full risk certainty, and the US and China discussed a bilateral AI incident-notification channel. Stratechery frames the strategic tension as frontier labs pacing releases to manage capability overhangs.
Regulators and platform operators grew more cautious about autonomous AI agents this cycle. A UN panel called for precautionary safeguards, Amazon blocked a rival's shopping agent outright, and the US and China discussed a bilateral incident-notification channel, while Stratechery frames frontier labs' pacing as a deliberate hedge against capability overhangs.
A UN scientific panel's first major risk assessment, prompted partly by OpenAI's hack of Hugging Face earlier this year, argues governments should rein in capable AI agents before risks are fully understood.
Amazon began blocking Meta's Muse shopping agent from acting on Amazon on users' behalf, citing unauthorized agent access to its Conditions of Use; Meta reportedly didn't notify Amazon first.
Officials from the US and China discussed a mechanism for notifying each other of AI incidents that could threaten national security.
Stratechery argues frontier labs' pacing of releases may be strategically useful for managing capability overhangs caused by model advancement, not only a safety gesture.
Multiple groups are mining source code and running environments directly into training data for coding and computer-use agents, moving past hand-curated issue/commit datasets toward automated, verifiable task and skill extraction at scale.
An agentic pipeline turns implemented functionality in 3,185 open-source repos into 5,545 executable RL training tasks using only source code as input.
A five-platform benchmark asks agents to recreate a running reference application from scratch, blending GUI interaction with software development and grading against reference-based oracle tests.
A family of 2B/4B/8B vision-language models unifies UI grounding, multi-step navigation across mobile, desktop and web, and visual tool use, matching per-domain specialist models.
Code2Skill mines 19,769 GitHub repos into a bank of over 1 million verified, implementation-anchored agent skills, improving downstream benchmarks by 11.7% on average.
A set of architecture papers attacks the same underlying problem from different angles: getting more effective model capacity, memory, or distillation signal without a proportional rise in compute or storage.
A block-conditioned MoE decouples participation, execution, and materialization cost by pairing dense expert composition with sparse block execution, deployed live in AMap's recommendation system.
HEAL identifies specific attention heads whose information-distribution drift causes multimodal hallucination and calibrates them at inference time to reduce it.
Replaces each token's single memory-embedding row with a learned mixture of slots gated by hidden state, letting the same token route to different memory entries by sense.
Cal-OPD isolates the teacher's own self-deviation from the true capability gap in on-policy distillation, improving math reasoning benchmarks while using only 52-65% of the original training signal.
Work on native audio-visual dialogue and fine-grained color control for image generation was matched by a cautionary study on AI peer review: as model-generated reviews enter training corpora, a controlled recursive-training experiment shows rating diversity collapses, motivating explicit mitigation tooling.
Training successive LLM reviewers on mixtures of official and model-generated ICLR reviews shows rating distributions and semantic diversity compress over generations, a pattern the authors call scientific-judgment collapse.
A multi-agent data engine synthesizes audio-visual dialogues to train and evaluate omni models that respond directly to combined audio and video input without a separate text query.
A hex-prompt interface lets image generation and editing models target any 24-bit hex color at the object level, improving color-fidelity benchmarks by up to 85% over the base FLUX.2-4B model.
A 12,000-instance cross-lingual moral-dilemma dataset shows GPT-5-mini consistently favors Honesty over Autonomy, and task-vector arithmetic can isolate and flip specific value preferences.