mlsanityData to 14 Oct 2026

Schema: Discovering Unknown Environments via Agentic Program Induction

arXiv 2609.39140 ↗First posted 30 Sep 2026cs.AIcited by 0

On the map: LLM agents and tool use → General LLM agent frameworks and benchmarks also fits: Agent skill learning and evolution

Abstract (arXiv)

Learning to complete tasks in unfamiliar environments with unknown rules remains a key challenge for LLM agents. Current LLM agents often record their discoveries in prose, which may not provide a compact, explicit account of how the environment works. Inspired by how scientists organize observations into testable, predictive theories, we introduce Schema, an agent harness that organizes learning and action through interactive program induction. The LLM agent decides what to investigate and how to act, expressing its evolving understanding of the environment as executable programs. The harness consists of a persistent program workspace and a small set of interfaces for checking these programs against the interaction history, planning within them, and executing plans under step-by-step verification. Schema raises ARC-AGI-3 RHAE from 58.7% to 99.2% with the same base model, solves 100% of the public DiG-bench games, and reaches the median performance of the top-50 human players on MazeBench. Extensive analysis shows the effectiveness of Schema in unknown mechanism discovery, and ablations confirm the contribution of each component.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in General LLM agent frameworks and benchmarks, one of 16 directions in LLM agents and tool use. Builds on 47 of its 75 references that are in the map. First posted on arXiv 30 Sep 2026.

Builds on

47 of 75 references.

Closest ancestors · 7

Same direction, newest first

  1. 2026Executable World Models for ARC-AGI-3 in the Era of Coding Agents6 May
  2. 2026ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence24 Mar
  3. 2025Test-Time Adaptation for LLM Agents via Environment Interaction6 Nov
  4. 2025Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents29 May
  5. 2025Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents20 Feb
  6. 2024τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains17 Jun
  7. 2024WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment19 Feb

Other references · 40

  1. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  2. 2023Reflexion: Language Agents with Verbal Reinforcement Learningcited by 4,276
  3. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  4. 2023SWE-bench: Can Language Models Resolve Real-World GitHub Issues?cited by 3,373
  5. 2023Voyager: An Open-Ended Embodied Agent with Large Language Modelscited by 2,191
  6. 2024SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineeringcited by 1,737
  7. 2023MemGPT: Towards LLMs as Operating Systemscited by 1,389
  8. 2019Mastering Atari, Go, Chess and Shogi by Planning with a Learned Modelcited by 1,353
  9. 2024OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environmentscited by 1,195
  10. 2026DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligencecited by 1,040
  11. 2026Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfacescited by 585
  12. 2024Executable Code Actions Elicit Better LLM Agentscited by 564
28 more
  1. 2024MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineeringcited by 469
  2. 2025LLMs Get Lost In Multi-Turn Conversationcited by 429
  3. 2025Agentic Context Engineering: Evolving Contexts for Self-Improving Language Modelscited by 398
  4. 2025Measuring AI Ability to Complete Long Software Taskscited by 156
  5. 2024BALROG: Benchmarking Agentic LLM and VLM Reasoning On Gamescited by 146
  6. 2023Hypothesis Search: Inductive Reasoning with Language Modelscited by 120
  7. 2025Competitive Programming with Large Reasoning Modelscited by 119
  8. 2023Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinementcited by 89
  9. 2021Human-Level Reinforcement Learning through Theory-Based Modeling, Exploration, and Planningcited by 43
  10. 2024Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Searchcited by 40
  11. 2018Investigating Human Priors for Playing Video Gamescited by 39
  12. 2025Why Do LLM Agents Fail in Exploring New Environments? A World-Modeling Perspectivecited by 28
  13. 2025PoE-World: Compositional World Modeling with Products of Programmatic Expertscited by 26
  14. 2024Doing Experiments and Revising Rules with Natural Language and Probabilistic Reasoningcited by 24
  15. 2025BoxingGym: Benchmarking Progress in Automated Experimental Design and Model Discoverycited by 18
  16. 2025Synthesizing world models for bilevel planningcited by 14
  17. 2026Prime Agent: A Self-Improving RLM Harnesscited by 13
  18. 2025TextQuests: How Good are LLMs at Text-Based Video Games?cited by 12
  19. 2025One Life to Learn: Inferring Symbolic World Models for Stochastic Environments from Unguided Explorationcited by 10
  20. 2025Benchmarking World-Model Learning with Environment-Level Queriescited by 10
  21. 2026PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoningcited by 10
  22. 2026OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration for ARC-AGI-3cited by 8
  23. 2026Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3cited by 7
  24. 2026Twin: Playing an Unknown Game with a Test-Time Digital Twincited by 4
  25. 2026Learning Abstractions for Hierarchical Planning in Program-Synthesis Agentscited by 3
  26. 2026NVIDIA-labs OO Agents: Native Python Object-Oriented Agentscited by 3
  27. 2026DiG-bench: Discovery in Gamescited by 3
  28. 2026Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025GroundAct: Can LLM Agents Ground Actions in Environmental States?
  2. 2026ProPlay: Procedural World Models for Self-Evolving LLM Agents
  3. 2026Action-Space Shaping for LLM Agents: Measuring and Mitigating Tool-Schema Bias
  4. 2026Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity
  5. 2026Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
  6. 2024AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning
  7. 2026Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning
  8. 2026ScrambleToolBench: Agents Search Exhaustively Even When Their Own Map Points to the Next Step