claude fable silently limits help on frontier ai development
anthropic's new safeguards for claude fable 5 silently reduce effectiveness on requests about building competing models, without notifying users.
topic
research
asr models struggle with bilingual code-switching
a new benchmark reveals how top automatic speech recognition systems handle mixed-language speech, with english segments causing the most errors.
gemini 3.5 live translate brings fluid speech-to-speech translation
google releases gemini 3.5 live translate, an audio model for near real-time speech-to-speech translation in over 70 languages with natural intonation.
syll: open-source agent for cross-surface automation
syll is an open-source, self-hosted agent that unifies api, cli, and gui control, letting users teach skills by demonstration and audit agent actions through logs and keyframes.
gemma 4 12b brings multimodal ai to laptops
google deepmind releases gemma 4 12b, an encoder-free multimodal model that runs locally on consumer laptops with 16gb ram.
spin: swarm control via tensorized policy coordination
a new framework uses tensor networks to make decentralized multi-agent swarm control scalable on edge devices.
ai tutoring trial boosts math scores in sierra leone
a randomized controlled trial in sierra leone found that ai-guided learning improved math scores by 0.258 standard deviations, with students showing more conceptual understanding over time.
recovering hidden components from unlabeled mixtures
new method identifies latent components in unlabeled data by exploiting marginal independence, without needing labels or clean samples.
persistent memory for continuous latent reasoning
a new method adds a gated concept stream to chain of continuous thought, letting models keep earlier facts during multi-step latent reasoning.
pathosage separates evidence for better pathology ai
a new framework reduces errors in pathology image analysis by independently checking conflicting evidence before making decisions.
omnimem compresses audio-visual llm memory for streaming video
a new framework reduces memory use in audio-visual large language models by separately managing visual and audio tokens and selecting only the most informative states.
why llms corrupt documents during delegation
a study finds that large language models silently corrupt documents over multiple edits, with smarter models fabricating plausible but false content.
how task diversity shapes in-context learning in transformers
a new analytical model shows that training task diversity, defined by non-overlapping low-dimensional subspaces, improves in-context learning by reducing interference and enabling better generalization.
why a market crash vanished when models changed
a bank run simulation that reliably crashed prices with one model stopped working when five different small models ran the same economy, revealing that emergent behavior is fragile and control requires authoring outcomes at settlement seams.
macarena benchmark tests ai agents on macos guis
macarena provides 421 tasks across 50 macos apps to evaluate computer-use agents on apple silicon, addressing gaps in existing benchmarks.
reusable safety adapters for fine-tuned language models
safegene introduces reusable adapter modules that restore safety alignment in fine-tuned open-weight llms without retraining from scratch.
diffusion model guides sudoku branch selection
a new method uses a diffusion model to rank candidate values in a symbolic sudoku solver, improving search efficiency while keeping correctness guarantees.
crowdmath dataset tracks collaborative math problem solving
a new dataset captures how groups of people work together to solve open math problems, showing the messy process of building proofs step by step.