verifiable-rewards

RL environments with verifiable rewards you can trust. Sandboxed coding and computer-use tasks, verifiers kept apart from the answer, a QA gate against reward hacking, and infra failures kept out of training. Works with verl, TRL, OpenRLHF and Axolotl.

Hillclimbing on verifiable rewards. Use Claude Code and Codex to autonomously search for python programs that optimize a score you define. 🏔️

DrugTargetWorld: a synthetic biobank for training and benchmarking AI scientists on end-to-end drug target discovery. Paper: arXiv:2610.09558

Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

Penalize the Path, Reward the Outcome — verifiable per-action penalties as a dense channel for deployable, sample-efficient agentic RL (GRPO). Paper: arXiv:2607.07435

[ NeurIPS MAR 2024 ] Official Codebase for "Dr-LLaVA: Visual Instruction Tuning with Symbolic Clinical Grounding"

Verifiable RL rewards for LLMs

Turn any real software into a replayable RL environment for training AI agents — deterministic replay, verifiable rewards, TRL & verifiers adapters.

An RL environment where a physics simulator (ngspice) is the verifiable reward — training a 3B LLM to design analog circuits. RLVR/GRPO, 2 decontaminated benchmarks, reward-hacking audit.

Agentic RL 最小实验台：把工具调用任务的成败做成可验证奖励（全部规则判定、不用模型当裁判），跑拒绝采样加 SFT 与 DPO。测试集达标率从 42.5% 提升至 62.5%

Outcome-verified agent trajectories, benchmarks, and RL environments — with a live leaderboard and a CI gate for your agents. Offline-first, MIT.

reinforcement learning with verifiable rewards using the online encyclopedia of integer sequences to generate diverse python programming tasks

Official implementation of Continual Reasoning Gym (CRG) and Continual Prompt Replay (CPR).

Zero-SFT reinforcement learning for Countdown arithmetic: GRPO/LoRA from a Qwen base model, exact rational verification, leakage-resistant splits, unbiased pass@k, curriculum experiments, OOD benchmarks, and reproducible evaluation reports.

Author verifiable eval records through a draft → review → revise → submit loop with server-enforced graders; compile to JSONL/CSV/Inspect/lm-eval via MCP. STDIO or Streamable HTTP.

Aerospace engineering LLM: open-weights base, continued pretraining, and RL with rewards from engineering tools (JSBSim, AVL, XFOIL, python-control, MIL-STD-1797A criteria). Ground tool; DO-330 path.

R1-style GRPO with verifiable rewards to improve 3D counting and spatial reasoning in Qwen2.5-VL (images and video)

GRPO reinforcement-learning fine-tuning, implemented from scratch in numpy (CPU-verified) + a TRL/Qwen GPU path on Modal. CLI-first, MLflow-tracked.

RLVR on FinQA: a verifiable numeric reward for financial reasoning, and a like-for-like online (GRPO) vs offline (DPO) comparison on the same signal. Leads with the reward-design and headroom findings.

Verifiable agent-planning task generators — tool ordering, budgets, loops, permissions, schedules; 100k-pool ready, exam harness included