Nanjing University's TAO-RL Fixes the Broken Training Loop Behind Agentic AI
TAO-RL fixes two root causes of unstable tool-use training in LLMs: bad data and timid exploration, beating existing methods across 7 benchmarks
20 stories
TAO-RL fixes two root causes of unstable tool-use training in LLMs: bad data and timid exploration, beating existing methods across 7 benchmarks
Ultralytics YOLO26 drops NMS post-processing, removes Distribution Focal Loss, and borrows an optimizer trick from LLM training to push real-time object detection to 57.5 mAP on COCO at under 12 ms latency.

Proof-Refactor is a four-phase agentic system that turns monolithic LLM-generated Lean proofs into readable, modular, library-quality artifacts
LAP proposes a missing third protocol layer that lets AI agents safely command real laboratory instruments, filling the gap left by MCP and A2A.
DOSS uses a learned classifier and LLM oversight to dynamically pick the right portfolio objective as market conditions shift, beating static strategies without the noise of regime detection.
Code-on-Graph turns KG facts into Python class interfaces, letting LLMs reason with code instead of bloated prompts — beating prior SOTA by up to 10.5%

IHO uses a masked diffusion model trained via preference optimization to jailbreak any LLM without white-box access, transferring across models and defenses without retraining
SkillPyramid organizes agent skills into a living hierarchy that composes, validates, and grows itself during task execution, boosting rewards by 38% across three benchmarks.
A new paper reveals that fine-tuning safety evaluations are fundamentally broken — and that capability must be measured alongside safety to draw any valid conclusions.
SHARS intercepts and resamples hallucinated text segments mid-generation, cutting factual errors without needing web search or external knowledge bases
AnchorMoE routes time series patches to specialized experts and explains every prediction as an exact additive sum over raw input segments — no post-hoc approximations needed.
TurtleAI exposes a critical blind spot in today's best vision-language models: even GPT-5 and o3 fail to translate geometric images into working Python code more than 60% of the time.

TSQAgent uses a three-role LLM framework with external tools to automatically identify and measure time series data quality, improving downstream forecasting and classification by selecting better training data
A new interpretability study reveals that fine-tuned LLM planners quietly build internal world models, and that training data diversity is the key to unlocking them.
VidMsg is a new benchmark exposing a critical blind spot in video-language models: understanding what a short clip is actually trying to say
NovelAPIBench reveals that usage examples beat signatures alone, and fine-tuning teaches models to use docs rather than memorize APIs
ChemCoTBench-V2 catches frontier LLMs getting the right answer for the wrong chemical reasons, using deterministic rule-based step verification across 5,620 samples
TTRL-CoCoV uses a model's own verification ability to fix label noise and prevent diversity collapse during label-free RL training, beating fully supervised methods on math benchmarks
A multi-institution research team introduces CRGC, a graph-based prompting framework that cuts LRM constraint violations by 39% without any model retraining
CauTion is a new academic framework that teaches LLMs when to stay quiet in causal discovery, using algorithm consensus to gate LLM input and calibrate trust without any labeled data.