mlsanityData to 14 Oct 2026

Learning to Discover at Test Time

arXiv 2601.16175 ↗First posted 22 Jan 2026cs.LG cs.AIcited by 115

Not on the map. Why?

Cited by 32 in LLM agents and tool use, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) ErdHos' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to 2× faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.

Key facts. Cited by 115 papers in the map (data through 14 Oct 2026). First posted on arXiv 22 Jan 2026.

Builds on

0 of 0 references.

No references here yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 115.

115 papers.

Sort:

By area:

  1. 2026Reinforcement Learning via Self-Distillationcited by 332
  2. 2026Meta-Harness: End-to-End Optimization of Model Harnessescited by 309
  3. 2026AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimizationcited by 60
  4. 2026EvoX: Meta-Evolution for Automated Discoverycited by 60
  5. 2026CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discoverycited by 52
  6. 2026Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perceptioncited by 50
  7. 2026Towards Autonomous Mathematics Researchcited by 40
  8. 2026Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpeningcited by 31
  9. 2026MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discoverycited by 28
  10. 2026DeltaEvolve: Accelerating Scientific Discovery through Momentum-Driven Evolutioncited by 26
  11. 2026K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Modelcited by 24
  12. 2026AVO: Agentic Variation Operators for Autonomous Evolutionary Searchcited by 24
38 more
  1. 2026SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?cited by 23
  2. 2026Structured Scaling of AI Discovery Across Diverse Scientific Domainscited by 21
  3. 2026SIA: Self Improving AI with Harness & Weight Updatescited by 21
  4. 2026RoboTTT: Context Scaling for Robot Policiescited by 19
  5. 2026Dream-RSI: Recursive Self-Improvement through Evolving Worldscited by 18
  6. 2026MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AIcited by 17
  7. 2026Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering Tasks with Generative Optimizationcited by 16
  8. 2026Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutionscited by 12
  9. 2026Harnessing Agentic Evolutioncited by 11
  10. 2026AI for Auto-Research: Roadmap & User Guidecited by 11
  11. 2026The Last AI Built by Humans: Toward Genuine Recursive Self-Improvementcited by 11
  12. 2026LLMs Improving LLMs: Agentic Discovery for Test-Time Scalingcited by 10
  13. 2026Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveriescited by 10
  14. 2026NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?cited by 10
  15. 2026Tool Verification for Test-Time Reinforcement Learningcited by 9
  16. 2026Autonomous Research Agents: A Survey of AI Scientists and the Verification Gapcited by 9
  17. 2026Meta-Reinforcement Learning with Self-Reflection for Agentic Searchcited by 8
  18. 2026Understanding the Challenges in Iterative Generative Optimization with LLMscited by 8
  19. 2026Escher-Loop: Mutual Evolution by Closed-Loop Self-Referential Optimizationcited by 8
  20. 2026Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineeringcited by 8
  21. 2026Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMscited by 7
  22. 2026Scaling Discovery through Test-Time Communicationcited by 7
  23. 2026Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimizationcited by 6
  24. 2026EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discoverycited by 6
  25. 2026Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Taskscited by 6
  26. 2026SwarmResearch: Orchestrating Coding Agents for Open-Ended Discoverycited by 6
  27. 2026Automated Discovery Has No Universally Superior Harnesscited by 6
  28. 2026Aster: Autonomous Scientific Discovery over 20x Faster Than Existing Methodscited by 5
  29. 2026CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programscited by 5
  30. 2026Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactionscited by 5
  31. 2026Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignmentcited by 4
  32. 2026Panini: Continual Learning in Token Space via Structured Memorycited by 4
  33. 2026Learning to Self-Evolvecited by 4
  34. 2026GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learningcited by 4
  35. 2026PACEvolve++: Improving Test-time Learning for Evolutionary Search Agentscited by 4
  36. 2026On the Generalization Gap in Self-Evolving Language Model Reasoningcited by 4
  37. 2026AgentStream: How Well Do Self-Evolving LLM Agents Perform Under Streaming Tasks?cited by 4
  38. 2026Policy of Thoughts: Scaling Test-Time Training for LLM Reasoning via Online Policy Evolutioncited by 3

50 most cited of 115 All 115 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
  2. 2025Scaling Test-Time Compute Without Verification or RL is Suboptimal
  3. 2025e3: Learning to Explore Enables Extrapolation of Test-Time Compute for LLMs
  4. 2026TTVS: Boosting Self-Exploring Reinforcement Learning via Test-time Variational Synthesis
  5. 2025Extending Test-Time Scaling: A 3D Perspective with Context, Batch, and Turn
  6. 2025TTRL: Test-Time Reinforcement Learning
  7. 2025Pushing Test-Time Scaling Limits of Deep Search with Asymmetric Verification
  8. 2025ThetaEvolve: Test-time Learning on Open Problems