AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning tasks at low cost.
hub Canonical reference
Learning to Discover at Test Time
Canonical reference. 100% of citing Pith papers cite this work as background.
24
Pith papers citing it
Background
100% of classified citations
abstract
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) Erd\H{o}s' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to $2\times$ 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.
hub tools
citation-role summary
background 10
citation-polarity summary
years
2026 24roles
background 10polarities
background 10representative citing papers
StarOR couples MCTS with GRPO-based test-time RL and unsupervised rewards to adapt optimization modeling policies instance-specifically, reporting SOTA results on five benchmarks with a 4B model.
Evolutionary coding agents achieve most benchmark gains through a small subset of edit types and by cycling previously deleted code lines rather than developing new algorithmic structures.
CODA re-expresses most non-attention Transformer computations as GEMM-plus-epilogue programs using a constrained set of composable primitives to keep intermediate results on-chip and cut global memory traffic.
EvoLib enables LLMs to accumulate, reuse, and evolve knowledge abstractions from inference trajectories at test time, yielding substantial gains on math reasoning, code generation, and agentic benchmarks without parameter updates or supervision.
AEvo introduces a meta-agent that edits the evolution procedure or agent context based on accumulated state, outperforming baselines by 26% relative improvement on agentic benchmarks and achieving SOTA on open-ended tasks.
Agentic-imodels evolves scikit-learn regressors via an autoresearch loop to jointly boost predictive performance and LLM-simulatability, improving downstream agentic data science tasks by up to 73% on the BLADE benchmark.
LLM-reinforced evolutionary search produces exact values Z(11,21,3,3)=116, Z(11,22,3,3)=121, Z(12,22,3,3)=132 and lower bounds for 41 additional Zarankiewicz numbers.
Meta-Harness discovers improved harness code for LLMs via agentic search over prior execution traces, yielding 7.7-point gains on text classification with 4x fewer tokens and 4.7-point gains on math reasoning across held-out models.
TTT-SCL dynamically generates test-aligned training sets for supervised causal learning using score-based functions and outperforms prior SCL and traditional causal discovery methods on benchmarks and real data.
MAP improves LLM agent reasoning by constructing a structured cognitive map of the environment before task execution, yielding performance gains on benchmarks like ARC-AGI-3 and superior training data via the new MAP-2K dataset.
UG-TTT adds epistemic uncertainty measured by adapter disagreement as an exploration bonus in RL for LLMs, raising maximum reward and diversity on scientific discovery benchmarks.
Tail-extrapolated estimators approximate best-of-N policy gradients from limited training rollouts by leveraging upper-tail reward statistics under structural assumptions.
SimpleTES scales test-time evaluation in LLMs to discover state-of-the-art solutions on 21 scientific problems across six domains, outperforming frontier models and optimization pipelines with examples like 2x faster LASSO and new Erdos constructions.
A jointly learned hierarchical index with cross-attention and residual quantization scales exact retrieval in foundational recommendation models, deployed at Meta with additional performance from test-time training on index nodes.
Frontier-Eng is a new benchmark for generative optimization in engineering where agents iteratively improve designs under fixed interaction budgets using executable verifiers, with top models like GPT 5.4 showing limited success.
TurboEvolve improves LLM program evolution by running parallel islands with LLM-generated diverse candidates that carry self-assigned weights, an adaptive scheduler, and clustered seed injection to reach stronger solutions at lower evaluation budgets.
GrandCode is the first AI system to consistently beat all human participants and place first in live Codeforces competitive programming contests.
Kernel-Smith combines evolutionary search with RL post-training to generate optimized GPU kernels, achieving SOTA speedups on KernelBench that beat Gemini-3.0-pro and Claude-4.6-opus on NVIDIA Triton and generalize to MetaX MACA.
Closed-loop self-evolution on LLMs improves reasoning on Knights and Knaves tasks but plateaus short of oracle-supervised levels, with multi-turn revision nearly matching it for large models.
PACEvolve++ uses a phase-adaptive reinforcement learning advisor to decouple hypothesis selection from execution in LLM-driven evolutionary search, delivering faster convergence than prior frameworks on load balancing, recommendation, and protein tasks.
Five improved inequalities were found with AI help: better Gaussian perimeter bounds for convex sets, sharper L2-L1 moments on the Hamming cube, a strengthened autoconvolution inequality, improved g-Sidon set bounds, and an optimal balanced Szarek inequality.
The paper delivers a stage-by-stage roadmap for AI in research, showing reliable assistance in retrieval and tool tasks but fragility in novelty and judgment, advocating human-governed collaboration.
citing papers explorer
-
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning tasks at low cost.
-
StarOR: Synergizing Tree Search and Test-Time Reinforcement Learning for Optimization Modeling
StarOR couples MCTS with GRPO-based test-time RL and unsupervised rewards to adapt optimization modeling policies instance-specifically, reporting SOTA results on five benchmarks with a 4B model.
-
What Do Evolutionary Coding Agents Evolve?
Evolutionary coding agents achieve most benchmark gains through a small subset of edit types and by cycling previously deleted code lines rather than developing new algorithmic structures.
-
CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs
CODA re-expresses most non-attention Transformer computations as GEMM-plus-epilogue programs using a constrained set of composable primitives to keep intermediate results on-chip and cut global memory traffic.
-
Test-Time Learning with an Evolving Library
EvoLib enables LLMs to accumulate, reuse, and evolve knowledge abstractions from inference trajectories at test time, yielding substantial gains on math reasoning, code generation, and agentic benchmarks without parameter updates or supervision.
-
Harnessing Agentic Evolution
AEvo introduces a meta-agent that edits the evolution procedure or agent context based on accumulated state, outperforming baselines by 26% relative improvement on agentic benchmarks and achieving SOTA on open-ended tasks.
-
Agentic-imodels: Evolving agentic interpretability tools via autoresearch
Agentic-imodels evolves scikit-learn regressors via an autoresearch loop to jointly boost predictive performance and LLM-simulatability, improving downstream agentic data science tasks by up to 73% on the BLADE benchmark.
-
New Bounds for Zarankiewicz Numbers via Reinforced LLM Evolutionary Search
LLM-reinforced evolutionary search produces exact values Z(11,21,3,3)=116, Z(11,22,3,3)=121, Z(12,22,3,3)=132 and lower bounds for 41 additional Zarankiewicz numbers.
-
Meta-Harness: End-to-End Optimization of Model Harnesses
Meta-Harness discovers improved harness code for LLMs via agentic search over prior execution traces, yielding 7.7-point gains on text classification with 4x fewer tokens and 4.7-point gains on math reasoning across held-out models.
-
Test Time Training for Supervised Causal Learning
TTT-SCL dynamically generates test-aligned training sets for supervised causal learning using score-based functions and outperforms prior SCL and traditional causal discovery methods on benchmarks and real data.
-
MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent Reasoning
MAP improves LLM agent reasoning by constructing a structured cognitive map of the environment before task execution, yielding performance gains on benchmarks like ARC-AGI-3 and superior training data via the new MAP-2K dataset.
-
Epistemic Uncertainty for Test-Time Discovery
UG-TTT adds epistemic uncertainty measured by adapter disagreement as an exploration bonus in RL for LLMs, raising maximum reward and diversity on scientific discovery benchmarks.
-
What should post-training optimize? A test-time scaling law perspective
Tail-extrapolated estimators approximate best-of-N policy gradients from limited training rollouts by leveraging upper-tail reward statistics under structural assumptions.
-
Evaluation-driven Scaling for Scientific Discovery
SimpleTES scales test-time evaluation in LLMs to discover state-of-the-art solutions on 21 scientific problems across six domains, outperforming frontier models and optimization pipelines with examples like 2x faster LASSO and new Erdos constructions.
-
Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation
A jointly learned hierarchical index with cross-attention and residual quantization scales exact retrieval in foundational recommendation models, deployed at Meta with additional performance from test-time training on index nodes.
-
Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering Tasks with Generative Optimization
Frontier-Eng is a new benchmark for generative optimization in engineering where agents iteratively improve designs under fixed interaction budgets using executable verifiers, with top models like GPT 5.4 showing limited success.
-
TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution
TurboEvolve improves LLM program evolution by running parallel islands with LLM-generated diverse candidates that carry self-assigned weights, an adaptive scheduler, and clustered seed injection to reach stronger solutions at lower evaluation budgets.
-
GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning
GrandCode is the first AI system to consistently beat all human participants and place first in live Codeforces competitive programming contests.
-
Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization
Kernel-Smith combines evolutionary search with RL post-training to generate optimized GPU kernels, achieving SOTA speedups on KernelBench that beat Gemini-3.0-pro and Claude-4.6-opus on NVIDIA Triton and generalize to MetaX MACA.
-
On the Generalization Gap in Self-Evolving Language Model Reasoning
Closed-loop self-evolution on LLMs improves reasoning on Knights and Knaves tasks but plateaus short of oracle-supervised levels, with multi-turn revision nearly matching it for large models.
-
PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents
PACEvolve++ uses a phase-adaptive reinforcement learning advisor to decouple hypothesis selection from execution in LLM-driven evolutionary search, delivering faster convergence than prior frameworks on load balancing, recommendation, and protein tasks.
-
Grokability in five inequalities
Five improved inequalities were found with AI help: better Gaussian perimeter bounds for convex sets, sharper L2-L1 moments on the Hamming cube, a strengthened autoconvolution inequality, improved g-Sidon set bounds, and an optimal balanced Szarek inequality.
-
AI for Auto-Research: Roadmap & User Guide
The paper delivers a stage-by-stage roadmap for AI in research, showing reliable assistance in retrieval and tool tasks but fragility in novelty and judgment, advocating human-governed collaboration.
- MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI