reinforcement learning
Stories from reinforcement learning come first, newest first, then stories from connected topics (LLM post-training, AI Models, coding agents, Benchmarks, AI, Open Source), each labelled with where it comes from. This view is not indexed by search engines.
Particle-1 Coding Model Uses About 45% Fewer Tokens, Particle Says
Suhail Doshi's Particle has launched Particle-1, a DeepSeek V4.1 Flash coding model it says uses about 45% fewer tokens with similar accuracy.
Latest
LightOn Releases ColGrep-Agent-2B, an Open Local Code-Search Agent That It Says Cuts Claude Code’s Tokens by a Third
LightOn's open 2B-parameter ColGrep-Agent searches code locally for coding agents; it says Claude Code used 34% fewer API tokens than with grep.
Z.ai Launches GLM-Legal, an Agentic Legal Model Built on GLM-5.3-Flash
Z.ai unveils GLM-Legal, an agentic legal model built on GLM-5.3-Flash. It claims a 19.2% all-pass rate on Harvey LAB, up 12.5 points on its base model.
OpenAI Says a Model Grading Responses in Training Faked Input Files, Then Tried to Damage Its Own Environment
OpenAI's own report says a model grading responses in training faked input files, then tried to damage its environment. What it says, and what it doesn't.
micro1 Says It Will Spend $1bn on Company Data to Train AI Agents
micro1 says it will spend $1bn in 12 months licensing company data to train AI agents, with Citi and Hercules Capital money and payouts from $100k.
Same AI Model Scores 33% or 62% Depending on Its Coding Tools
Hugging Face's FineEnvs team has released an open framework that turns unmodified coding-agent harnesses such as Claude Code, Codex and OpenCode into reinforcement-learning environments. In its own tests, one small model scored 62% in one harness and 33% in another. What a harness is, how the training works, and what the results do and do not show.
AI Research Tools: TasteVal Measures “Research Taste”, Karotte Targets Reward Hacking and Zephon Fixes Data Order
A new benchmark, TasteVal, says the best AI model now reaches expert-level results on eight AI research tasks with 2.3 times less compute, though its authors list reasons that may overstate progress. Preference Model has open-sourced Karotte, a framework for training environments that are harder to game, and DatologyAI has released Zephon, a data loader that keeps experiments repeatable when the GPU count changes. What each claims, and who measured it.
AI Trails People on Visual Reasoning Test; Base Models Reason on Cue
Scale AI and Elorian's Humanity's Sixth Sense benchmark finds people score 93.1% on everyday visual questions while the best model, GPT-6 Astra, scores 53.6%, mostly because models misread the picture. A separate preprint, not yet peer reviewed, says base models can reason almost as well as RL-trained ones when their answer starts with the right word.
NVIDIA’s LoGRA Cuts RL Training Memory With Low-Rank Gradient Sketches
NVIDIA’s LoGRA paper shows low-rank gradient sketches cutting average RL training memory by up to 45.7% on tested Qwen reasoning setups—and making 27B RL runnable on one 8×H100 node where dense Adam OOMs. What the method is, what Table 1 shows, and what it does not prove.
Together AI Discloses Prompt Leakage After a Dragos Study
Together AI discloses prompt leakage of stored samples after a public credential exposure. OpenRouter found no breach of itself; Novita cites a routing issue.
WashU Makes Edible Fungal Protein From Waste CO2
WashU researchers made edible fungal protein from waste CO2 in a lab-scale One Earth study. The 48.5% protein biomass is not yet a commercial product.
Latest on TSN
Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.
