EPOCH

The EPOCH route still has unresolved issues: the computational overhead of the coverage objective on larger-scale problems, the quality stability of teacher LLM insights, and others still require follow-up work to answer.

The EPOCH framework takes instance coverage directly as the objective of evolutionary search and, through experience preference optimization, enables the generator to internalize the reflection signal.

EPOCH optimizes coverage simultaneously from three aspects: parent selection prioritizes individuals that can fill coverage blind spots, operator allocation dynamically adjusts strategies according to coverage gaps, and the stagnation recovery mechanism seeks new directions that cover untouched instances when at a local optimum.

EPOCH's search process tends to return a complementary ensemble of heuristics rather than a single optimal solution.

EPOCH introduces a reflection self-distillation (RSD) term that explicitly increases the likelihood of the generator autonomously producing improved versions.

The paper validated on multiple combinatorial optimization problems and multiple open-source LLM backbones, and the heuristic ensembles generated by EPOCH surpassed inference-only and fine-tuning baselines in performance and could generalize across problem scales.

Heuristic combinations trained on smaller-scale instances remain effective when transferred to larger-scale instances.

EPOCH brings substantive progress on ten concrete discovery problems, including improved executable constructions, optimized algorithms, counterexamples, and proof-backed results, spanning both the mathematical and computational domains.

The paper argues that evidence governance is a necessary step toward trustworthy AI research agents.

EPOCH implements an evidence-governed discovery loop composed of five core components: explicit task contracts, typed memory, active falsification, admission checks, and independent replication.

EPOCH decouples search optimization from evidence adjudication, and counters false-positive findings that only hold under specific conditions through active falsification and independent replication.

Typed memory attaches an evidence-type label to every stored piece of information, distinguishing different levels such as numerical test passed, formal verification passed, and validated on training set only.

EPOCH achieves state-of-the-art overall performance on the AlgoTune benchmark, with an average normalized score of 0.65, surpassing the strongest baseline's 0.53.

EPOCH achieves the highest average score of 0.57 on the internal Math14 test suite.

EPOCH leads on the descriptive composite metric on AgentHPO.

A latest arXiv paper proposes EPOCH — an architecture centered on evidence-governed principles, with the goal of enabling AI research agents to produce more trustworthy scientific discoveries.