Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.35873

More Programs or More Rolls? Separating Coverage from Specialization in LLM Harnesses

Ziyang Xu·Haitian Zhong·Hao Zhou·Hao Qin·Chenhan Jin·Te Qi·Shengze Xu·Tieyong Zeng
View original source ↗Read source document ↗Export metadata ↓

Abstract

Automated generation of LLM harnesses promises to improve inference through task specialization. Yet additional answer coverage can arise from repeated execution of the same program, making specialization difficult to identify. We introduce a controlled evaluation that separates answer coverage, repeatable task advantages, and gains from pre-execution selection. On 386 MATH-500 tasks, we compare eight generated harnesses plus a baseline with nine byte-identical baseline copies, using three executions per member. Identical programs yield 2.16 percentage points of repeat-averaged oracle headroom. Generated programs exhibit substantially more repeatable score patterns, but these chiefly reveal persistent weaknesses: losses relative to the baseline persist across all three repeats on 100 tasks, while persistent wins occur on only one task and are sensitive to answer extraction. The frozen selector gains 0.00 percentage points, and both populations reach 98.70% oracle coverage at 27 harness executions. Stable complementarity remains unresolved at three repeats. Supporting BIRD traces locate failures in mechanism implementation, activation, and output validity. Together, these findings establish why coverage and repeatability alone cannot justify claims of useful specialization. They motivate an evaluation standard for harness diversity: task advantages should persist across executions, guide usable decisions, and improve on additional fixed-program executions under matched inference budgets.

Explore related subjects

cs.AIcs.LGcs.SE

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ziyang Xu, Haitian Zhong, Hao Zhou, Hao Qin, Chenhan Jin, Te Qi, Shengze Xu, Tieyong Zeng. 2026-09-26. More Programs or More Rolls? Separating Coverage from Specialization in LLM Harnesses. https://arxiv.org/abs/2609.35873

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Generalising from Self-Produced Data: Model Training Beyond Human Constraints

Current large language models (LLMs) are constrained by human-derived training data and limited by a single level of abstraction that impedes definitive truth judgments. This paper introduces a novel framework in which AI models autonomously generate and validate new knowledge through direct interaction with their environment. Central to this approach is an unbounded, ungamable numeric reward - such as annexed disk space or follower count - that guides learning without requiring human benchmarks. AI agents iteratively generate strategies and executable code to maximize this metric, with successful outcomes forming the basis for self-retraining and incremental generalisation. To mitigate model collapse and the warm start problem, the framework emphasizes empirical validation over textual similarity and supports fine-tuning via GRPO. The system architecture employs modular agents for environment analysis, strategy generation, and code synthesis, enabling scalable experimentation. This work outlines a pathway toward self-improving AI systems capable of advancing beyond human-imposed constraints toward autonomous general intelligence.

cs.AI↗
arXiv2026

Hierarchical Reasoning Model

Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT) techniques, which suffer from brittle task decomposition, extensive data requirements, and high latency. Inspired by the hierarchical and multi-timescale processing in the human brain, we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains significant computational depth while maintaining both training stability and efficiency. HRM executes sequential reasoning tasks in a single forward pass without explicit supervision of the intermediate process, through two interdependent recurrent modules: a high-level module responsible for slow, abstract planning, and a low-level module handling rapid, detailed computations. With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples. The model operates without pre-training or CoT data, yet achieves nearly perfect performance on challenging tasks including complex Sudoku puzzles and optimal path finding in large mazes. Furthermore, HRM outperforms much larger models with significantly longer context windows on the Abstraction and Reasoning Corpus (ARC), a key benchmark for measuring artificial general intelligence capabilities. These results underscore HRM's potential as a transformative advancement toward universal computation and general-purpose reasoning systems.

cs.AI↗
arXiv2026

A memory-based active inference model of DishBrain-like adaptive behaviour

Recent and rapid advances in artificial intelligence (AI) make it increasingly important to understand the foundations of adaptive behaviour in autonomous agents, especially for building safe and efficient systems. While artificial neural networks have dominated the development of AI, recent work has begun to explore living biological neuronal networks as an alternative substrate for computation. These systems promise remarkable data and sample efficiency and rich dynamics, and may also inspire explainable and biologically plausible models. Here, we develop an experiment-informed active inference framework to model decision-making in closed-loop agents that mirror experimental setups using biological neurons. Using a generative model whose dimensions are matched to an experiment protocol, we systematically compare three decision-making schemes within this common generative model. Under matched episode counts (i.e. total data available for learning) to the in-vitro experiment, our simulations show that agents with short memory horizons reach a level of performance close to that of mouse and human cortical cultures (DishBrain platform), whereas longer memory horizons depart from it substantially. Increasing the planning horizon, by contrast, confers no comparable benefit. Because all model parameters are explicit, we can also track the quantities in our generative model that accompany this improvement, such as the risk term and the entropy of the transition and state-action mappings. Together, these results illustrate how active inference offers a formal language for comparing decision-making schemes in similar closed-loop control environments.

cs.AI↗