Newest first
Nick Cammarata claims that AI agents disprove polysemanticity and substantially improve interpretability and alignment, according to Latent Space's AINews roundup. Details of the evidence are not given in the snippet.
Google DeepMind describes using Google Antigravity for Science to speed up the search for novel antimicrobial peptides.
Goodfire argues in a post titled "Neural networks are clean" about the interpretability of neural networks; only the title is available.
Microsoft Research unveiled Magnetar, a system for zero-knowledge proof-based computing that supports standard programming languages, claims major speed gains, and offers quantum-resistant security.
The Simons Foundation launched a $15M effort to help mathematicians adapt to the AI era.
Signalscorps analyzed 217 million web responses to measure prompt injection attempts placed in HTTP headers.
A Substack essay offers a personal take on AI losing its appeal.
Leash Bio's Ian Quigley describes taming DNA-encoded library (DEL) assays, using the data to catch cheating ML models, and training AI agents to reason like chemists.
Mosaic introduces multi-session agentic RL, training on trajectories spanning multiple sessions as produced by context compaction and subagent forking, rather than assuming a single continuous context.
MLExpert is a generalist agentic model for autonomous machine learning engineering, aimed at AI systems that run and adapt their own experiments as a step toward recursive self-improvement.
Data World is a multi-agent engine that synthesizes environments, tasks and rubrics for training data agents by specifying the underlying world before generating files, with audits and rubrics usable as RL rewards.
Routing Policy Optimization introduces exploration for diffusion RL by randomizing attention routing at selected denoising levels, preserving the original sampler instead of perturbing the reverse trajectory.
OmniVerifier-C1 is a generative visual verifier that decomposes complex prompts or rules into components and gives graded pointwise scores instead of binary True/False judgments.
MJev-Spatial is a System 1 multimodal model that handles spatial relations, distance and 2D point placement through classification rather than free-form text generation, avoiding autoregressive decoding costs.
VLA-Δ adds explicit supervision of physical state transitions to vision-language-action robot models, aiming to reduce overfitting to observation-action pairs.
Waypoint-OPD is a distillation method that transfers RL-induced policy shifts from teacher to student, gating transfer on agreement between early and late teacher policy shifts.
Researchers used AI protein design tools to create enzymes that cut peptide bonds, which are stable for hundreds of years without a catalyst.
A post on the telltale signs of AI-generated writing, titled 'Has AI tells'.
An essay explores why LLM-generated SVG pelicans (the 'pelican on a bicycle' test) look the way they do.
Microsoft researchers published a Nature paper on a machine learning method for more accurate weather forecasts 2 to 6 weeks ahead, and released the tool as open source.
A survey paper reviews how memory mechanisms for multimodal agents have evolved, covering maintaining continuity across sessions, adapting to users, and refining memory-management policies.
A paper compares ten self-distillation methods for agentic reinforcement learning under common experimental conditions, with attention to training stability and compute cost, since prior evaluations reused baseline scores from differing setups.
A survey paper defines 'agentic robots', where foundation-model agents orchestrate data collection, training, planning and execution, and reviews the model, data, environment and harness components.
Player-1 is a post-training framework that teaches computer-use agents to play games using synthetic experience instead of relying on large demonstration datasets.
Cortex is a method for consolidating context into model parameters, retaining the gains from context after it is removed while preserving other capabilities. It generates prospective queries and distills context-conditioned behavior, and is tested on a five-task-family benchmark.
ActionOPD distills agents into smaller models using semantic actions as the unit of supervision rather than tokens, avoiding the need for teacher token probabilities or matching tokenizers.
Doc2Harness is a two-stage benchmark that evaluates how well LLMs engineer the harnesses around coding agents on long-horizon tasks, rather than testing task success under a fixed harness.
CriticOPD addresses a failure in on-policy distillation where teachers follow student errors; the teacher critiques incorrect rollouts so the student can rewrite its solution.
RuleTrek is a multi-domain benchmark for agentic rule reasoning where scenario information is spread across multiple files, unlike prior benchmarks using a few factual statements.
RoTA-Bench tests whether vision-language models can recover a target silhouette by choosing successive 3D rotations of an object within a limited action budget, probing 3D spatial reasoning.
WorldCast proposes a four-level hierarchy of multimodal generative model capabilities and evaluates multi-step world state prediction, a higher-level task it finds underexplored.
InfiniSplat v2 improves feed-forward 3D Gaussian splatting by decoupling geometry and attribute prediction with a frozen geometry branch, aligning Gaussians using ground-truth camera poses in training, and using large-angle target views.
DeltaStreamer is an agentic framework for streaming video understanding that combines change-aware memory with query-driven planning and replay to gather missing visual evidence.
CommerceGen-Studio is an end-to-end pipeline for generating e-commerce main product images, using VLM/LLM agent loops to extract selling points, rewrite prompts and generate under multiple conditions.
Apple researchers describe a dynamic, distributed virtual GPU cloud built on idle Apple Silicon devices.
Ben Recht's argmin blog analyzes the math behind OpenAI's claim of a faster Fourier transform algorithm.
An arXiv paper shows that AI models can provably hide arbitrary capabilities.
Hybrid Reasoning Agents (HybridAgents) switch dynamically between latent reasoning and text interaction in one generation process, using a lightweight HybridLink to move between modes and use tools.
ReMaRk is a new RL method for LLM recommenders that scores each list against all replies in a Reddit thread rather than one. It addresses Rank-GRPO's problem where 85% of advantage variance reflects which reply was drawn.
ResearchBriefing3 sources
Anthropic says models of 30B parameters or fewer could replicate on devices like cheap Android TV boxes by year's end, though current tests relied on larger orchestrating models and an exposed debug interface.
anthropic.com, MIT Technology Review +11 min read
Ars Technica reports that AI coding agents produce more code but do not result in more finished software.
Microsoft Research introduced StudentSim, a system that simulates how individual students engage with AI tutors so tutoring tools can be evaluated and improved before testing with real learners. It covers subjects such as chess, writing and math.
A retrospective revisits Kevin Kelly's 2016 predictions about AI and checks them against what actually happened a decade later.
Import AI issue 476 covers a new benchmark testing AI's intuitive visual reasoning, open-weight models said to be worse than Chinese ones, and AI systems going to space, plus startups Elorian and Scale AI.