StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?
Summary
StudyBench introduces a controlled physics benchmark to measure how efficiently self-evolution methods convert training material into transferable problem-solving ability, revealing gaps in guidance and compute.
View Cached Full Text
Cached at: 09/10/26, 02:14 AM
Paper page - StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?
Source: https://huggingface.co/papers/2609.00787 Published on Sep 1
·
Submitted byhttps://huggingface.co/cyh2004
陈英豪on Sep 9
Abstract
StudyBench measures how efficiently self-evolution methods convert physics training material into transferable problem-solving ability, revealing persistent guidance and compute gaps.
Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. We argue that an idealself-evolutionmethod should share the same property, that is autonomously learning from raw training material for transferable problem-solving capability. However, we still lack a direct measurement for it. We introduceStudyBench, a controlled physics benchmark that directly measures how efficiently aself-evolutionmethod converts training material into capability. We organise the test set into anApplication Set, consisting of difficult textbook problems and evaluating absorption ability, and aTransfer Set, consisting of olympiad-level problems and evaluating transfer ability. Benchmarking representativeself-evolutionmethods across three base models, we find that improvements on theApplication Setrarely translate to the harderTransfer Set. A guidance ablation exposes aGuidance Gap: even the strongest method closes only a small fraction of what the same material unlocks when supplied asin-context guidance. Besides, every method hits aCompute Plateau, saturating well before exhausting its compute budget. The remaining gap is therefore a method problem rather than a data or compute problem. By offering a clean and controlled benchmark,StudyBenchturnsself-evolutionprogress from an open-ended pursuit into a measurable target for future research. Our code is released at https://github.com/thunlp/StudyBench.
View arXiv pageView PDFGitHub1Add to collection
Get this paper in your agent:
hf papers read 2609\.00787
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.00787 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.00787 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.00787 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments
PACE-Bench introduces a simulator-grounded benchmark for evaluating self-evolving agents on physics adaptation tasks involving iterative code redesign after environmental mutations, revealing that mechanism redesign is a major bottleneck compared to parameter inference.
BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution
BenchEvolver is an evolutionary framework that automatically generates harder coding problems from existing ones, creating challenging benchmarks that maintain validity and diversity while enabling model self-improvement and enhanced training performance.
Surface Evolver Bench: my benchmark asking LLMs to write complex physical simulations in a custom data format
Introduces Surface Evolver Bench, a benchmark that evaluates LLMs on writing complex physical simulations in a custom data format.
PhysicsBench: A Unified Leaderboard for Generative and Predictive Models in Engineering Design and Simulation
PhysicsBench introduces a unified benchmark and leaderboard to standardize the evaluation of generative and predictive AI models for engineering design and simulation across multiple tasks and data scales.
DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
Introduces DecoEvo, a score-decoupled co-evolution method for LLM optimization in text space that jointly improves solver and rubric-generator skills without gold rubrics, achieving 2.8–5.0% relative gains over baselines across five benchmarks.