Amazon AGI paper proposes AutoGym for agent RL environments

eelvis9/28/2026, 20:22:54

A paper from Amazon AGI proposes AutoGym for generating reinforcement-learning environments for agents. AutoGym writes the task, the executable environment, and the verifier together, starting from a small domain seed or past model trajectories.

The paper is arXiv:2609.22592. A Chat with Paper page is also available.

#autogym #amazon-agi #reinforcement-learning #ai-agents

View original post ↗

Previous articleOpenRouter says six decision models are now on the market, two of them free Next articleClaude Sonnet 5.5 is now live in Command Code

Related AI news

CMU paper trains a 4B proposer to edit agent harnesses without changing the solver

The harness-learning paper uses reinforcement learning so a proposer revises harness code from a task, the current harness, and an execution report. A trained 4B proposer beats its 35B teacher on single-step Reasoning Gym revisions, and HotpotQA training transfers to MuSiQue and 2WikiMultihopQA.

eelvis•10/3/2026, 17:20:53•

#harness-learning #cmu #ai-agents #reinforcement-learning #reasoning-gym

Viktor offers $100 credit trial for Slack agent evaluation assistant

Viktor provides an AI employee in Slack that analyzes AI agent harness evaluation logs, traces regressions to specific changes, and suggests rollbacks, offering a free trial with $100 in credits and no credit card required.

eelvis•10/7/2026, 09:03:53•

#viktor #ai-agents #evaluations #slack

NVIDIA paper details VERA framework for co-evolving agent harnesses and models

An NVIDIA research paper presents VERA, a system that converts benchmark trajectories into over 9,000 restartable sandboxes with rubric scoring while co-evolving agent harnesses alongside model weights. The team open-sourced the environment corpus, reporting that a 27B agent scored 71.6 on AutoCoWorkBench to top Claude Opus 4.8.

eelvis•10/7/2026, 08:30:48•

#nvidia #ai-agents #benchmarks #open-source

Harness-Aware Distillation trains small model agents to 63.4% success on ALFWorld

A research paper presents Harness-Aware Distillation, a training framework that teaches small language model agents to leverage harness information by contrasting teacher decisions. The resulting student model reached 63.4% on unseen ALFWorld tasks, outperforming its 8B teacher.

eelvis•10/6/2026, 09:57:17•

#model-distillation #ai-agents #alfworld #llm-research

VeriHarness turns a base model into a verifier for long-horizon agent tasks

A Google paper, VeriHarness, uses the same base model to check both disputed and unanimous agent claims. With evidence-backed revision, it reports gains of 6.2 points with Gemini 3.5 Flash and 6.4 points with Claude Opus 4.8 over a single rollout.

eelvis•10/4/2026, 06:53:11•

#veriharness #google #ai-agents #gemini #claude

Reset history AI News FAQ Usage limits RSS feeds API MCP Skill

News is compiled from public information. The data is for reference only, and this site has no affiliation with OpenAI.

Data & original project: Did Codex Reset · Licoy