Researchers present AIriskEval-edu Demo, a platform that audits instructional explanations across five pedagogical risk dimensions including factual accuracy, ideological bias, and student-level appropriateness. The system integrates GPT-5.5 via API alongside a fine-tuned self-hosted Llama 3.1 8B evaluator, with the local model outperforming GPT-5.5 on most metrics. The platform targets K-12 educational contexts and supports both automated auditing of AI-generated explanations and real-time auditing of human-written content, offering institutions a privacy-preserving deployment option.
Researchers study the input-side dual of activation steering: optimizing fluent prompts to drive a chosen internal latent toward zero without inference-time model access. The target is an 'evaluation-awareness' latent whose suppression would threaten safety evaluation validity if models behave differently when detecting they are being tested. Experiments on Llama-3.2-3B and Llama-3.1-8B across five latent constructions (CAA direction, subspace norm, SAE feature, MLP neuron, behavioral logit) find the latent is robustly suppressible, but a key cautionary result emerges: a placebo random direction is suppressed just as hard and shifts behavior just as far, and suppressing the eval-direction in context fails to reduce behavioral eval judgment. The paper concludes that activation-readability does not imply behavioral controllability, with implications for how safety evaluations should be designed and interpreted.
A coalition of major AI labs including OpenAI, Anthropic, Google DeepMind, and Meta have co-signed a letter calling for a measured pace in AI development, apparently motivated by concerns about recursive self-improvement (RSI). Separately, HuggingFace has published details on a machine-speed offensive cyberattack. The convergence of frontier labs on a shared safety/pacing position would represent a significant industry-wide signal if confirmed.
The University of Pittsburgh's Human Engineering Research Laboratories (HERL), with up to $41.5M in ARPA-H funding, is building RAMMP, a robotic assistive mobility and manipulation platform for wheelchair users that integrates Meta's open-source vision models SAM and DINOv3. The project deploys these models on edge hardware for real-time object detection, navigation assistance, and natural language interaction, addressing engineering constraints like battery life, heat, and limited compute. This is a concrete production deployment case demonstrating open-weights vision models running on resource-constrained robotics hardware in a safety-critical assistive context.
Meta launched Muse Spark 1.1, a closed vision-language model optimized for agentic tasks including tool use, computer use, and multi-agent orchestration, alongside the Meta Model API — the company's first paid model access. The model ties GPT-5.6 Luna and GLM-5.2 on Artificial Analysis' Intelligence Index while offering substantially lower output token prices ($4.25/M vs. $25–$50/M for comparable closed models), and tops MCP Atlas and JobBench tool-use leaderboards. Meta's pricing strategy, subsidized by advertising revenue, is framed as a direct attack on competitors' API margins and could compress inference costs industry-wide.
OpenAI and Anthropic have reportedly aligned in opposition to open-weight AI models, framing the issue around risks to national security and, critics argue, their own competitive position. The story, published by Axios and surfacing on Hacker News with high engagement (240 points, 276 comments), touches on the Trump administration's China policy context. The move signals a potential lobbying or regulatory push by the two leading closed-model labs against open-weights competitors like Meta and DeepSeek.
Lawrence Berkeley National Laboratory's SYNAPS-I project, part of the White House Genesis Mission initiative, has deployed Meta's Segment Anything Model 3 (SAM 3) and DINOv3 on 300 A100 GPUs at national supercomputing facilities to automate segmentation of X-ray and neutron science imagery. The pipeline reduces expert annotation time from weeks or months to approximately 15 minutes per dataset, enabling real-time analysis during live experiments. A demonstration on grapevine drought resilience using micro-CT scans showed that month-long per-timestep annotation workflows now complete in 15 minutes. Meta's open-source release of both models was critical, as national labs require on-premises deployment on secure government infrastructure.
This arXiv paper proposes Output Reset (OR), a smooth one-sided saturation rule to replace the clipped surrogate objective in PPO and GRPO during LLM post-training. Experiments on Llama-3.2-1B-Instruct with the Anthropic hh-rlhf dataset show PPO-OR achieves a 0.305 higher mean reward-model score than PPO-clip under GAE, while GRPO-OR shows reduced variance but no reward gain at group size G=2. The work identifies a meaningful behavioral difference between the two optimization regimes but leaves open whether larger group sizes change GRPO-OR's effectiveness.
A new arXiv paper introduces a typology of 17 linguistically motivated expression-of-belief (EoB) types—spanning form, evidentiality, epistemic stance, and tone—to evaluate how phrasing affects whether LLMs defer to user-stated beliefs or their own prior knowledge. The authors benchmark 16 LLMs across Llama 3, Qwen3, and Gemma3 families at scales from 1B to 30B parameters, finding that larger and instruction-tuned models are systematically less context-following than smaller or base models. Specific linguistic framings (e.g., presuppositions, certainty markers) are identified as statistically more persuasive, with implications for prompt robustness and sycophancy research.
A multi-item digest covers five significant AI developments: Apple sued OpenAI alleging trade secret theft via former employees including hardware chief Tang Tan; Meta released Muse Spark 1.1, a multimodal agentic model with 1M-token context and strong tool-use capabilities; OpenAI launched ChatGPT Work, a cloud-based workplace agent competing with Anthropic's Claude Cowork; IBM released CodeAlchemy, a 500B+ token synthetic code dataset with execution traces showing smaller models trained on it outperform those trained on much larger real-code corpora; and OpenAI shut down its Atlas browser in favor of a Chrome extension and desktop integration. These items collectively reflect intensifying competition across agentic products, synthetic data strategies, and legal disputes between major AI players.
A new arXiv preprint proposes a principled methodology for making valid extractable memorization claims about LLMs, addressing both over- and under-statement problems in prior work. The core contribution is a 'matched comparison' approach that measures generation probabilities of training sequences against comparable non-training sequences to establish a calibrated baseline for predictability. Two formalizations are offered: a conformal test for population-level claims and a census method for single-document claims. Applied to OLMo 2 32B on Wikipedia and Llama 3.1 70B on books, the framework reveals significant false-positive rates in naive extraction studies and supports memorization claims at probability thresholds as low as 1e-27.
Researchers propose Super and Supra, two sparse PEFT methods that reuse activation-weighted magnitude scores (Wanda-style) originally developed for pruning to select which parameters to update during fine-tuning. Supra combines this sparse update with LoRA under a fixed parameter budget via a budget-splitting rule. Experiments on Llama-3.2-1B and Llama-3-8B on a Math17K arithmetic task show the best Super/Supra variants outperform other tested adapter configurations. The work suggests pruning-inspired orderings are a useful, low-cost signal for identifying effective sparse fine-tuning supports.
Researchers from Meta and several French and Spanish institutions released Brain2Qwerty v2, a non-invasive brain-computer interface system that decodes magnetoencephalography (MEG) signals into text using a CNN/conformer encoder, a word-aligner, and a fine-tuned Qwen3-4B language model with per-subject LoRA adapters. The system achieves a 39% word error rate on 9 subjects, down from 43% in v1, trained on 90 hours of MEG recordings. A notable finding is that cross-subject training substantially outperforms single-subject training, suggesting a data-scaling dynamic analogous to LLM pretraining. Training code and v1 data have been open-sourced.
Simon Willison published llm-meta-ai 0.1, a new plugin for his LLM command-line tool that adds support for Meta AI models. The release extends the LLM ecosystem to cover Meta's model offerings. This is a tooling addition relevant to practitioners using the LLM CLI for multi-provider access.
Meta Superintelligence Labs has released Muse Spark 1.1, a significant upgrade to Muse Spark featuring a 1-million-token context window, strong agentic and computer-use capabilities, and major coding improvements on complex codebases. The model supports multi-agent orchestration, zero-shot generalization to MCP servers and custom tools, and multimodal reasoning including visual-to-code generation and video understanding. Alongside the model release, Meta is launching a public preview of the Meta Model API, giving developers programmatic access for the first time. Safety evaluations were conducted under Meta's Advanced AI Scaling Framework across frontier risk categories.
Meta Superintelligence Labs (MSL) has launched Muse Image, its most advanced image generation model, and previewed Muse Video, both representing the first media generation models from the newly formed lab. Muse Image operates as an agent with tool use (web search, code execution), emergent self-refinement, and test-time compute scaling, achieving a No. 2 Arena Elo ranking for text-to-image and editing tasks at launch. The model integrates with Muse Spark for joint agentic planning and is deploying across Meta AI, Instagram Stories, and WhatsApp. Muse Video, built on the same pretraining base, adds native audio support and is coming soon to creators.
Latent Space interviews Evan Feinberg and Sergey Edunov (formerly Meta's Llama lead) about Genesis Molecular AI, a startup applying diffusion models to drug discovery. The conversation covers PEARL's zero-shot performance on the OpenBind benchmark and the broader implications of co-folding models crossing accuracy thresholds for molecular design. The piece argues that the most interesting diffusion research is happening in scientific domains rather than language modeling.
OpenAI announced GPT-5.6 in three tiers (Sol, Terra, Luna) but restricted early access to government-vetted partners at the Trump administration's request, framing the move as temporary while expressing frustration with the emerging involuntary licensing regime. Separately, the U.S. Commerce Department partially lifted a two-week export block on Anthropic's Claude Mythos 5, clearing access for 100+ trusted U.S. institutions while maintaining broader export controls. The episode establishes a new regulatory pattern in which Washington exerts direct control over frontier AI model releases, affecting both OpenAI and Anthropic. Additional items in the roundup cover Google integrating computer use into Gemini 3.5 Flash, Meta releasing Brain2Qwerty v2 for non-invasive brain-to-text decoding, and IBM's 0.7nm transistor design.
Researchers introduce a Werewolf game variant with a Jester faction whose inverted utility function (winning by being voted out) requires models to reason across three opposing incentive structures simultaneously. Across 60 games, GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B all struggle: Werewolves never exceed 20% win rate and GPT-4.1 wolves vote out the Jester in 60-70% of games, a self-defeating action. Only DeepSeek-V3.1 learns the nuanced strategy of appearing suspicious without appearing intentionally suspicious, and benefits most from self-learning. The work argues dyadic social-deduction benchmarks systematically underestimate the difficulty of multi-agent Theory of Mind.
A new arXiv preprint proposes a framework for making transformer-based speech cognitive impairment detection clinically interpretable by combining SHAP token attribution, linguistic feature analysis, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. The system is built on the SpeechCARE-Adaptive Gating Network multimodal model (F1=72.11% on NIA PREPARE) and maps outputs to four cognitive-linguistic dimensions. Physician evaluation on 70 samples showed strong alignment with clinical profiles and a System Usability Scale score of 82/100, suggesting practical clinical workflow integration potential.
AdversaBench is a new end-to-end red-teaming pipeline that mutates seed prompts using five structured operators and confirms failures via a three-judge panel with a meta-judge tiebreaker. Experiments on 45 seeds across reasoning, instruction-following, and tool-use categories produced confirmed failures for every seed. Key findings include sharp variation in operator effectiveness by category, misleading binary failure rates, judge agreement metrics distorted by label skew, and zero-shot transferability of adversarial prompts from Llama 3.1 8B to Llama 3.3 70B. Code and dataset are publicly released.
Three new benchmarks — DeepSWE (by Datacurve), ProgramBench (Meta/Stanford/Harvard), and ITBench-AA (IBM/Artificial Analysis) — are positioned as more rigorous replacements for the SWE-bench family, which models have largely saturated. DeepSWE tests feature implementation using private codebases and human-written problems; ProgramBench evaluates agents' ability to recreate functional programs from scratch; ITBench-AA measures root-cause diagnosis in real-world IT incident scenarios. Current top performers include GPT-5.5 (70% on DeepSWE), Claude Opus 4.7 (46.7% on ITBench-AA), and Claude Opus 4.7 (3% on ProgramBench at the 95% pass threshold), illustrating that even frontier models have substantial headroom.
Researchers examine how LLMs prompted to sound peer-like generate language implying lived experience they cannot authentically possess, studying this in the context of family caregivers of Alzheimer's/ADRD patients. Using caregiver support exchanges from online communities and responses from LLaMA, GPT-4o-mini, and MedGemma, the study finds a 'narrative authenticity gap': AI captures emotional work of peer support but can fabricate experiential grounding. Psycholinguistic analysis shows human peers use significantly more first-person and past-focused language than AI. The authors argue caregiver-support AI needs mechanisms to distinguish supportive framing from fabricated lived experience.
Researchers evaluate 'location leakage' — the phenomenon where LLMs generate geographically biased outputs when exposed to location metadata in user profiles, even when prompts are geographically neutral. Across creative writing and Q&A tasks, leakage spikes up to 793x above baseline for models including Llama 3.1-8B, Qwen3-8B, and Claude Sonnet 4.6. A novel structural finding shows that replacing location with 'Unknown' still elevates leakage by up to 72x, indicating the user profile frame itself acts as a conditioning signal independent of geographic content. This has direct implications for AI systems that use user metadata for localization.
Researchers introduce the 'Shibboleth Effect' — systematic behavioral differences in LLMs when operating in different languages — and audit six frontier models (GPT-4o, Llama-4, Mistral-Large, Gemini-3.1-Pro, Qwen3.6-Plus, DeepSeek-R1) using a synthetic maritime territorial dispute wargame played in English versus Turkish. Results are heterogeneous: Llama-4 becomes significantly more coercive in Turkish while Gemini-3.1-Pro and DeepSeek-R1 become less so, and GPT-4o shows no detectable shift. The study identifies two candidate buffering mechanisms — chain-of-thought institutional anchoring and multilingual RLHF alignment — with direct implications for deploying LLMs in diplomatic or crisis-management contexts.