# Witness: Discovery, Deciphering, and Epiphany in Interactive Puzzle Environments

arXiv: https://arxiv.org/abs/2609.32208 · HTML: https://mlsanity.com/paper/2609.32208/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 18 of its 26 references that are in the map.
- First posted on arXiv 26 Sep 2026.

## Builds on (most cited first, 18 of 18)

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code](https://mlsanity.com/paper/2403.07974/) (2024)
- [SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning](https://mlsanity.com/paper/2506.24119/) (2025)
- [Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles](https://mlsanity.com/paper/2505.19914/) (2025)
- [lmgame-Bench: How Good are LLMs at Playing Games?](https://mlsanity.com/paper/2505.15146/) (2025)
- [Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play](https://mlsanity.com/paper/2509.25541/) (2025)
- [ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence](https://mlsanity.com/paper/2603.24621/) (2026)
- [Game-RL: Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning](https://mlsanity.com/paper/2505.13886/) (2025)
- [Measuring General Intelligence with Generated Games](https://mlsanity.com/paper/2505.07215/) (2025)
- [PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors](https://mlsanity.com/paper/2507.15550/) (2025)
- [Prime Agent: A Self-Improving RLM Harness](https://mlsanity.com/paper/2608.23552/) (2026)
- [CipherBank: Exploring the Boundary of LLM Reasoning Capabilities through Cryptography Challenges](https://mlsanity.com/paper/2504.19093/) (2025)
- [WILT: A Multi-Turn, Memorization-Robust Inductive Logic Benchmark for LLMs](https://mlsanity.com/paper/2410.10998/) (2024)
- [Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models](https://mlsanity.com/paper/2604.02485/) (2026)
- [PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts](https://mlsanity.com/paper/2506.06211/) (2025)
- [HardcoreLogic: Challenging Large Reasoning Models with Long-tail Logic Puzzle Games](https://mlsanity.com/paper/2510.12563/) (2025)
- [CausalGame: Benchmarking Causal Thinking of LLM Agents in Games](https://mlsanity.com/paper/2607.04293/) (2026)
- [DiG-bench: Discovery in Games](https://mlsanity.com/paper/2608.12593/) (2026)

## Abstract (arXiv)

Automated science needs agents that can work out the rules of an unfamiliar environment by interacting with it. Interactive rule-discovery puzzles offer a controlled setting for studying this ability: an agent infers hidden rules through experimentation and uses what it has inferred to reach a stated goal. We ask what limits current language models on these puzzles and whether reinforcement learning (RL) improves performance on rules held out from training. To study both, we introduce WITNESS, a 2D grid-based puzzle environment with ground-truth ASCII observations and controlled access to rules. An agentic pipeline generates games for WitnessGym, the RL training suite, and WitnessBench, comprising public validation and private test games. The validation set separately tests new compositions of trained rule primitives and primitives absent from training. Under a shared harness, the best of 18 frontier proprietary and open-weight models solves only 24% of private test level slots, with scores sensitive to the observation interface and agent configuration. Providing ground-truth rules raises Opus-5's validation RHAE-L5 (relative human action efficiency over the first five levels) from 59.9 to 97.8, whereas a 27B open-weight model gains only 2.1 points and remains limited even with the rules provided. RL on WitnessGym raises the 27B model's private test RHAE-L5 from 2.1 to 5.4 and yields a mean gain of 4.1 points on four external discovery benchmarks. Together, these results point to rule acquisition as a major difficulty for frontier models like Opus-5 while smaller models further struggle on rule-based execution, and indicate that RL on hidden-rule puzzles transfers to broader rules and real-world tasks beyond training. Benchmark is available at: https://witnessbench.ai