
Microsoft Finds One Signal Boosts AI Code-Fixing Agents by 44 Points
Microsoft and Georgia Tech researchers inject perfect information into SWE agents to rank which signals actually move the needle on bug fixing.
A focused feed of models, agents, research and open-source releases for people building with AI.

Microsoft and Georgia Tech researchers inject perfect information into SWE agents to rank which signals actually move the needle on bug fixing.

An independent open-source take on Anthropic's internal Scion gives teams a Git-compatible server built for swarms of coding agents landing changes in parallel.

Sakana AI's QD-Rec applies Quality-Diversity evolutionary search to LLM-powered recommendations, growing archives of varied, well-justified picks instead of ranked lists.

DeepSeek open-sourced DeepInfer and deepseek-recipe, the production stack powering DeepSeek-V4.1-Flash, bringing their internal serving optimizations and API parity to anyone.

A preliminary Vals AI test shows an open-weight DeepSeek model outperforming OpenAI's gated cyber tier on a multi-step network attack scenario at a fraction of the cost.

A new framework retrieves full scientific papers by decomposing them into aspect-specific queries and iteratively expanding the search as a tree.

An 8-step distilled MiniMax H3 that turns image, video, and voice references into clips with synced audio now runs natively in ComfyUI.

Vals AI shipped v2.2 of its GDP-weighted agent benchmark, tightening four component tests and reshuffling the frontier model leaderboard.

Ant Group quietly dropped an MIT-licensed 0.5B TTS model that ties Fish Audio for #1 pronunciation accuracy and leads all open weights in voice cloning.

SGLang's new plugin framework lets NVIDIA and AMD ship specialized low-latency kernels, cutting DeepSeek-V4.1-Flash decoder time by up to 1.56x.

A new post-hoc calibration method called Truth AnChoring maps raw LLM uncertainty scores to probabilities grounded in actual correctness.

Artificial Analysis launched independent leaderboards for OpenAI's two health conversation benchmarks, surfacing sharp tradeoffs between answer length, cost, and clinical quality.

OpenAI's newest models jump from single-digit to near-perfect accuracy on COBOL migration and show massive gains across 20 programming languages.

A 27B fine-tune of Qwen3.8 built for everyday messages beats much larger models at sounding like an actual human being.

Anthropic's new OSS Scanner uses Claude Mythos to send unreviewed vulnerability reports with reproducers and patches directly to open-source maintainers.

Prime Intellect ships a new default RL loss built for async training, plus a 1.15x throughput bump over torchtitan on Qwen3-235B.

Meta's 30B open-weight agentic model runs on a single consumer GPU at 4-bit, beating Gemma4-31B and Qwen3.6-27B on tool use and planning.

Black Forest Labs and fal shipped a post-trained FLUX 3 variant that renders a 10-second clip in 12 seconds, 22x faster than Seedance 2.5 at matched quality.

Artificial Analysis launched AA-PersonalityBench, scoring 36 models on 30 personality traits, 8 indexes and 21 programmatic communication metrics using judged pairwise comparisons.

Meta's Prometheus supercluster hits 85% goodput and 47K GPU-hour mean time between interruptions across hundreds of thousands of GPUs training Muse Wonder.

A new meta-evaluation framework shows frontier AI agents cave to confident but wrong user hints, with task scores falling as much as 46.7 percent.

Together AI launched a post-training service that lets you write your own RL or SFT loop in Python while they manage GPU orchestration, rollout engines, and weight sync.

A new xLSTM variant corrects one scalar instead of a whole vector per memory update, cutting interference and improving long context performance.

Gemini now decides which video segments to actually look at, cutting long-form video analysis costs by up to 66% and token usage by up to 88%.
No stories match these filters.