Source-linked AI summary
HeavySkill: Heavy Thinking as the Inner Skill in Agentic Harness
Jianing Wang, Linsen Guo, Zhengyu Chen, Qi Guo, Hongyu Zang, Wenjie Shi, Haoxiang Ma, Xiangyu Xi, Xiaoyu Li, Wei Wang, Xunliang Cai
TL;DR
Agentic harnesses achieve strong complex-task performance, but the mechanism behind that success remains unclear. HeavySkill abstracts the harness into parallel reasoning followed by sequential deliberation, finding consistent gains over Best-of-N and performance approaching Pass@K for high-reasoning models.
Problem
The mechanism driving agentic harness performance remains unclear behind complex orchestration systems.
Method
HeavySkill decomposes inference into parallel reasoning and sequential deliberation, with memory caching to organize and revisit reasoning trajectories.
Results
Heavy thinking consistently outperforms traditional Best-of-N strategies across tested tasks, with high-reasoning models approaching the theoretical Pass@K limit.
Takeaways & Limitations
Reinforcement learning can substantially improve reasoning capabilities by scaling the depth and width of heavy thinking as a learnable skill.
Takeaways & Limitations
Benefits are less pronounced or occasionally negative on preference-oriented tasks such as Arena-Hard.
Abstract
from arXiv · showhide
Recent advances in agentic harness with orchestration frameworks that coordinate multiple agents with memory, skills, and tool use have achieved remarkable success in complex reasoning tasks. However, the underlying mechanism that truly drives performance remains obscured behind intricate system designs. In this paper, we propose HeavySkill, a perspective that views heavy thinking not only as a minimal execution unit in orchestration harness but also as an inner skill internalized within the model's parameters that drives the orchestrator to solve complex tasks. We identify this skill as a two-stage pipeline, i.e., parallel reasoning then summarization, which can operate beneath any agentic harness. We present a systematic empirical study of HeavySkill across diverse domains. Our results show that this inner skill consistently outperforms traditional Best-of-N (BoN) strategies; notably, stronger LLMs can even approach Pass@N performance. Crucially, we demonstrate that the depth and width of heavy thinking, as a learnable skill, can be further scaled via reinforcement learning, offering a promising path toward self-evolving LLMs that internalize complex reasoning without relying on brittle orchestration layers.
1. Introduction
The introduction frames HeavySkill as a two-stage heavy-thinking skill for agentic harnesses: parallel reasoning followed by sequential deliberation or summarization. It presents a training-free framework, systematic evaluation, and analyses identifying trajectory quality and diversity as key performance factors.
- Heavy thinking pipeline: Heavy thinking decomposes into parallel reasoning that generates independent trajectories and sequential deliberation that aggregates them into a final answer.This two-stage pattern is presented as the core mechanism underlying orchestrated heavy thinking.
- Framework: HEAVYSKILL provides a training-free framework with parallel reasoning, sequential deliberation, and a memory cache for organizing reasoning trajectories.The cache supports iterative handling of generated trajectories within the framework.
- Empirical findings: Ablations find that trajectory quality and diversity in parallel reasoning are two key contributors to performance.The analysis also reports that sequential deliberation largely depends on the employed model’s general capability.
- Contributions: The paper contributes a comprehensive empirical study of heavy thinking across diverse model scales and task domains, including its effectiveness and limitations.This study is positioned alongside the proposed training-free framework and systematic analysis.
2. Methodology
HeavySkill decomposes inference into parallel reasoning followed by sequential deliberation, connected through a serialized memory cache and optionally repeated iterative refinement. The same protocol is expressed as a readable skill that orchestrators can execute autonomously and port across compatible harnesses.
- Inference Pipeline: Heavy thinking uses two phases: parallel reasoning generates independent trajectories, followed by sequential deliberation that aggregates them into summary content.The phases use separate LLMs, πθ for trajectories and πϕ for summarization.
- Memory Cache: A serialized memory cache stores candidate trajectories between phases, while pruning and shuffling address context-length limits and inference robustness.Complete trajectories may contain extensive internal thinking and answer content, so serializing all of them can exceed the model’s maximum length.
- Iterative Deliberation: Iterative deliberation modifies the cache by concatenating previous sequential-deliberation content, enabling repeated refinement across iterations.The procedure defines iterations t ∈ {2, · · · , N} and uses concatenation to update the cache.
- Readable Skill: The HEAVYSKILL document is a structured natural-language executable specification with four components for activating and directing heavy thinking.The supplied passage identifies the four-component structure, while the surrounding methodology describes the skill as a readable document for orchestrators.
- From Workflow to Skill: Unlike workflow mode’s external Python orchestration, skill mode lets the LLM orchestrator read the document and autonomously execute the protocol using serialized context.The orchestrator spawns parallel agents, collects outputs, and routes them through its context window.
- Portability and Generality: Because HEAVYSKILL is plain text without framework-specific dependencies, it functions under Claude Code and custom orchestration harnesses without modification.This portability supports the paper’s thesis that heavy thinking is an inner skill rather than an artifact of one system design.
3. Experiments
Experiments across STEM and general reasoning tasks show that heavy thinking improves over average parallel trajectories, often approaches or exceeds raw sampling potential, and benefits most tasks with objective correctness criteria. Its gains are weaker on subjective preference tasks, while summary potential indicates additional re-reasoning capacity.
- Experimental scope: Heavy thinking is evaluated across STEM benchmarks including AIME25, BeyondAIME, HMMT25-Feb, and GPQA-Diamond.The experiments compare HM@4 and HP@4 with Mean@K, Pass@K, and Majority Voting.
- STEM results: HM@4 consistently surpasses M@K across models and STEM benchmarks, showing gains from parallel reasoning followed by sequential deliberation.The improvement is reported across all evaluated STEM models and benchmarks.
- STEM results: Increasing parallel width and deliberation depth continues improving performance rather than causing a plateau, supporting heavy thinking as test-time scaling.The framework uses K ∈{8, 16} parallel trajectories and K(1) = 4 generated summaries in the main experiments.
- Scaling and potential: Heavy thinking frequently exceeds Majority Voting and can approach Pass@K on frontier models, while HP@4 can surpass P@K when the deliberation model is sufficiently capable.The reported pattern suggests sequential synthesis can re-reason beyond simply selecting an existing trajectory.
- General reasoning results: 69.3% is the HM@4 result for R1-Distill-Qwen-32B on IFEval, up from 35.7% M@K, while GPT-OSS-20B reaches 85.5% HM@4 on LiveCodeBench.These gains occur on tasks with clear logical or programmatic constraints.
- General reasoning results: On Arena-Hard, HM@4 gains are more marginal or occasionally slightly negative, indicating weaker benefits for subjective, preference-oriented alignment.The passage contrasts these results with correctness-oriented tasks where sequential deliberation is more effective.
4. Further Analysis
Further analysis shows that HeavySkill corrects difficult low-confidence trajectories, preserves high accuracy on easier queries, generalizes across model pairings, and scales with iterative deliberation. It also improves tool-using reasoning over majority voting, although additional iterations can degrade HP@K.
- Distributional Analysis: Over 500 queries with parallel pass rates below 0.5 are successfully rectified through heavy deliberation, demonstrating corrective potential where heuristic voting struggles.Approximately 1,400 queries remain unresolved in this low-pass-rate cohort.
- Distributional Analysis: Above a parallel pass rate of 0.5, heavy thinking maintains a success rate exceeding 98%, while approximately 30 queries experience negligible degradation.The reported losses are small relative to the overall dataset gains.
- Cross-Model Generalization: HM@K consistently outperforms baseline M@K across AIME25 and HMMT25-Feb regardless of whether sequential deliberation uses R1-Distill-Qwen-7B, R1-Distill-Qwen3-8B, or Qwen2.5-32B-Instruct.The results support compatibility across architectures and cross-model collaboration.
- Iterative Deliberation: Increasing deliberation iterations consistently improves HM@K, indicating intrinsic scaling, but significantly degrades HP@K, suggesting later steps may suffer interference from information.Experiments use K = K(1) = · · · = K(N) = 8 and N = 4.
- Tool-Using Reasoning: 90.0% accuracy is achieved by GPT-OSS-20B with HM@4 on AIME25, versus 83.3% with V@4, and HM@4 consistently surpasses voting across tested models and datasets.The tool-using trajectories receive execution feedback from a Python interpreter.
5. Related Work
Recent work frames parallel thinking and post-training with test-time scaling as important approaches to improving LLM reasoning. Existing methods often rely on independent trajectories, fixed synchronization, handcrafted heuristics, or extended reasoning chains, whereas heavy thinking offers a more flexible implementation of test-time scaling.
- Parallel Thinking: Parallel thinking has emerged as an efficient test-time scaling technique for reasoning.Current approaches commonly spawn independent trajectories with end-stage aggregation or synchronize thoughts at rigid, predefined intervals.
- Parallel Thinking: Tree of Thoughts enables granular parallelization but remains dependent on handcrafted heuristics and external verifiers.These limitations also characterize approaches that synchronize thoughts at fixed intervals.
- Heavy Thinking: Heavy thinking provides a more flexible approach to test-time scaling, as demonstrated in Gemini, Kimi K2, and PaCoRe.The paper focuses on implementing the heavy thinking pattern and analyzing its effectiveness.
- Post-training and Test-time Scaling: Post-training with test-time scaling has become a powerful and versatile technique for LLM reasoning through systems such as OpenAI’s o1, DeepSeek R1, and Gemini.These approaches aim to develop self-correction, reflection, critique, and verification, typically by extending reasoning chains through long CoT.
6. Conclusion
The paper frames heavy thinking as a readable test-time reasoning skill for agentic harnesses, structured around parallel reasoning and sequential deliberation. Evaluations find that it outperforms traditional Best-of-N strategies and can approach Pass@K in models with high intrinsic reasoning potential.
- Framework: Heavy thinking is presented as a novel test-time strategy that enhances LLM reasoning and distills the workflow into a readable skill for agentic harnesses.The framework centers on parallel reasoning and sequential deliberation.
- Empirical findings: Heavy thinking consistently outperforms traditional Best-of-N strategies, including voting methods.The conclusion attributes this to how test-time computation translates into task performance.
- Empirical findings: In models with high intrinsic reasoning potential, heavy-thinking performance can approach the theoretical Pass@K limit.The paper also conducts extensive detailed analysis of heavy thinking’s effects.
A. Impact of Parallel Trajectories
The study evaluates how selecting among parallel reasoning trajectories affects performance on AIME25 and HMMT25-Feb. Accuracy rises as the number of trajectories increases across all selection methods, while Max-Diversity performs comparably to Random sampling.
- Experimental setup: Using R1-Distill-Qwen3-8B, the study evaluates Random, Max-Diversity, Max-Length, and Max-Answer-Num selection over 256 parallel trajectories on AIME25 and HMMT25-Feb.The experiment investigates the impact of permutations in parallel reasoning.
- Accuracy trends: Accuracy consistently increases for all selection methods as the number of parallel trajectories K grows.The result indicates that increasing inference budget through parallelization benefits performance regardless of the selection heuristic.
- Selection strategies: Max-Diversity achieves performance comparable to Random sampling.This suggests that Max-Diversity does not provide a clear performance advantage over Random in the reported comparison.
B. Advancing Heavy Thinking via RLVR
This section investigates whether RLVR can improve heavy-thinking trajectories and expand models’ reasoning capabilities. Experiments reuse parallel reasoning trajectories and show early gains, including an approximately 10% improvement in HM@4, alongside stability differences between cache widths.
- Motivation: Heavy thinking can outperform heavy majority under specific scenarios, motivating RLVR on heavy-thinking trajectories.The observation is framed as a question of whether RLVR can further elevate the model’s reasoning upper bound.
- Experimental Setup: The study reuses parallel reasoning trajectories, selecting queries with pass rates in [0, 0.625] and sampling K ∈ {8, 16} trajectories as serialized memory caches.Training uses VeRL with GSPO and R1-Distill-Qwen-7B as the backbone model.
- Results: Approximately 10%: HM@4 improves further during the initial 100 training steps, while both training and test performance grow consistently.The reported divergence in stability occurs after 100 steps, with significant entropy collapse observed for K = 16.
C. Prompt and SKill for Heavy Thinking
This section defines HeavySkill as a reasoning-amplification skill for complex tasks, combining parallel reasoning with sequential deliberation. It also specifies the prompt and implementation artifacts used to support this process.
- Implementation: The agentic-harness skill file is documented across Figures 8, 9, and 10.These figures present parts I, II, and III of the heavy-thinking skill file.
- Prompt: The heavy-thinking prompt has multiple thinkers independently develop thought processes while incorporating one another’s progress toward the final answer.The prompt begins by assigning the model the role of a strong reasoner.
- Prompt: The memory-cache prompt is presented in Figure 7 as the serialized-memory implementation of heavy thinking.The passage identifies Figure 7 specifically as the prompt for memory cache.
- HeavySkill overview: HeavySkill decomposes complex problem-solving into parallel reasoning followed by sequential deliberation that synthesizes trajectories into a superior final answer.The skill is intended for cases where a single chain-of-thought may be insufficient.