> TL;DR: HeavySkill, proposed by Meituan's LongCat team, splits complex reasoning into two stages: "parallel thinking → sequential discussion." Experiments show that a weaker model acting as the "discussion moderator," combined with several mid-tier "thinker" models, can outperform a single strong model working alone. Kimi K2 Thinking reaches 100% accuracy on AIME25 with HeavySkill. Most strikingly, the deliberation stage can not only select the correct answer but synthesize a brand-new correct answer out of flawed traces—HP@K can even exceed the Pass@K ceiling.
Key Points
- Problem with Best-of-N (BoN): Majority voting has a blind spot—truth can sit in the minority. If 3 of 16 runs answer correctly (42) and 13 answer wrongly (37), BoN picks 37.
- Method: Stage 1 generates K fully independent reasoning traces (temperature 1.0, top-p 0.95, top-k 10). Stage 2 has a deliberator model read all traces, critically assess each, and re-derive if all thinkers are wrong. Deliberation is synthesis, not voting—it can produce a new correct answer absent from all traces.
- Results: Kimi K2 Thinking hits 100% on AIME25 (HM@4). On IMO problems, GLM 4.6 reaches 86.0% HP@4 vs. a 74.5% Pass@K ceiling—HP@K exceeding Pass@K.
- Deliberator need not be strong: Qwen2.5-32B-Instruct scores only 12.8% solo on AIME25 but as deliberator lifts a 7B thinker to 50–60% (K=8). Deliberation requires synthesis and critical evaluation, not peak reasoning.
- Trace selection: Consensus-based selection (Max-Answer-Num) works best; preferring the longest traces is worst—"verbosity ≠ quality."
- Iterative deliberation: HM@K rises over rounds, but HP@K may fall due to error contamination; 2–3 rounds recommended.
- RLVR training: VeRL + GSPO on R1-Distill-Qwen-7B lifts HM@4 from ~35% to ~45% (K=8); K=16 hits ~48% but shows entropy collapse after 100 steps. Context length is the bottleneck.
- Deliverable: A plain-text skill file that plugs into Claude Code or any agentic harness with zero code changes.
- Paper: https://arxiv.org/abs/2605.02396
- Code: https://github.com/wjn1996/HeavySkill
- Skill file:
skill/heavyskill.md(plain-text, plug-and-play; supports vLLM, DeepSeek, Together AI, OpenRouter, Ollama)
Ablation Highlights
| Strategy (selecting K from 256) | Effect | |---|---| | Max-Answer-Num (consensus) | Best | | Random / Max-Diversity | Similar, mediocre | | Max-Length | Worst — verbosity ≠ quality |
Limitations
1. Compute cost: K inference passes + 1 deliberation = K+1× single-run cost. 2. Context length: K=16 deliberation can exceed limits, destabilizing RLVR training. 3. Not a panacea: Simple tasks don't need it. 4. HP@K degrades with iterations due to error accumulation.
Key Numbers
| Metric | Value | |---|---| | Paper | arXiv:2605.02396 | | Authors | 11 (Jianing Wang et al., Meituan LongCat team) | | Best AIME25 | Kimi K2 Thinking: 100% (K=16) | | HP@K > Pass@K | GLM 4.6 IMO: 86.0% vs 74.5% | | Deliberator effect | Qwen2.5-32B: 12.8% solo → 50–60% as deliberator | | RLVR gain | ~10% absolute (stable at K=8) | | License | Apache-2.0 |
Resources
Takeaway
HeavySkill's core claim: let AI agents first reason in parallel, independently, then use one round of reasoning to synthesize all traces. The deeper implication—how multiple AIs collaborate may matter more than how smart a single AI is. Reasoning capability may lie more in the organization of inference than in parameter count. If a plain-text skill file can bring any model close to Pass@N performance, the bottleneck may not be training new models at all.
---
*Reference: Wang et al., "HeavySkill: Heavy Thinking as the Inner Skill in Agentic Harness", arXiv:2605.02396, 2026. Also see https://blog.qiaomu.ai/heavyskill-heavy-thinking-inner-skill-agentic*