HeavySkill Deep Dive: Why AI 'Group Discussion' Beats Majority Voting

> TL;DR: HeavySkill from Meituan's LongCat team splits complex reasoning into two stages — *parallel thinking* then *sequential discussion*. Experiments show a weak model acting as "discussion moderator" plus several mid-tier "thinkers" can outperform a single strong model. Kimi K2 Thinking hits 100% on AIME25 with HeavySkill. Most strikingly, the discussion stage can synthesize brand-new correct answers — HP@K even exceeds the Pass@K ceiling.

1. The Problem: Majority Voting Can Be Wrong

The mainstream strategy for complex AI reasoning is Best-of-N (BoN) — run the model N times and vote. Its fatal blind spot: the truth may sit with the minority. If 16 attempts yield the correct answer 3 times and a wrong answer 13 times, BoN picks the wrong one.

Humans solve hard problems differently: first think independently from different angles, then sit down and discuss — question each other, combine strengths, converge on the best answer. HeavySkill ports this human logic to AI, echoing Feynman's learning technique: explaining exposes gaps in understanding.

2. The Two-Stage Pipeline

Stage 1: Parallel Reasoning

• Generate K fully independent reasoning traces, each starting from scratch
• Diversity encouraged (algebraic vs. geometric approaches)
• Sampling: temperature 1.0, top-p 0.95, top-k 10
• This is brainstorming, not retrying 16 times

Stage 2: Sequential Deliberation

A deliberator model (same or different) reads all traces and: 1. Classifies the problem type to set analysis depth 2. Critically evaluates each trace — "don't blindly follow the majority" 3. Re-derives from scratch if all thinkers are wrong 4. Keeps language and format consistent

Discussion is synthesis, not filtering. It frequently synthesizes a correct answer absent from every trace.

Iteration

For very hard problems, up to 4 rounds: HM@K rises with rounds, but HP@K may drop as later rounds are contaminated by earlier errors. Practical cap: 2–3 rounds.

3. Key Results

AIME25

| Model | Single | Pass@8 | Voting@8 | HM@4 | HP@4 | |---|---|---|---|---|---| | Kimi K2 Thinking | 95.4% | 100% | 96.7% | 100% | 100% | | GPT-5 Thinking | 92.5% | 100% | 96.7% | 96.7% | 96.7% | | DeepSeek R1-0528 | 87.1% | 96.7% | 90.0% | 93.3% | 93.3% | | R1-Distill-Qwen3-8B | 76.7% | 90.0% | 83.3% | 85.8% | 90.0% |

At K=16 gains grow: GPT-5 Thinking HM@4 hits 99.2%, DeepSeek R1-0528 96.7%.

BeyondAIME (harder)

| Model | Pass@8 | Heavy HP@4 | |---|---|---| | Kimi K2 Thinking | 87.0% | 84.0% | | GPT-5 Thinking | 86.0% | 83.0% | | DeepSeek R1-0528 | 84.0% | 77.0% |

Coding (LiveCodeBench)

| Model | Pass@K | HP@4 | |---|---|---| | GPT-OSS-20B | 89.0% | 85.5% | | GLM 4.6 | 90.3% | 87.9% |

HP@K > Pass@K (IMO)

| Model | Pass@K | HP@4 | |---|---|---| | GLM 4.6 | 74.5% | 86.0% | | GPT-OSS-20B | 65.8% | 84.5% |

The discussion stage can stitch together fragments of multiple wrong answers into a new correct answer — impossible for BoN.

4. Ablations

Trace selection (pick K from 256)

• Max-Answer-Num (consensus): best
• Random ≈ Max-Diversity
• Max-Length: worst — verbosity ≠ quality

Does the deliberator need to be strong? No.

With fixed thinkers (R1-Distill-Qwen-7B):

| Deliberator | AIME25 K=8 | HMMT25 K=8 | |---|---|---| | R1-Distill-Qwen-7B | 36.25% | 26.66% | | R1-Distill-Qwen3-8B | 56.66% | 60.00% | | Qwen2.5-32B-Instruct | 50.00% | 40.00% |

Qwen2.5-32B-Instruct scores only 12.8% solo on AIME25, yet reaches 50–60% as a deliberator. Deliberation needs synthesis, critique, and summarization — not top-tier reasoning.

Iteration (K=8, 4 rounds)

HM@K improves steadily (e.g., R1-Distill-Qwen-7B on HMMT25: ~32% → ~40%; DeepSeek R1-0528: ~88% → ~93%), but HP@K can decline.

5. RLVR Self-Improvement

• Framework: VeRL + GSPO; model: R1-Distill-Qwen-7B
• K=8: HM@4 from ~35% to ~45%, stable
• K=16: up to ~48%, but entropy collapse after 100 steps
• Bottleneck: sequence length limits in the discussion stage

6. Skill Format: Plug-and-Play Reasoning

HeavySkill is packaged as a plain-text skill file (skill/heavyskill.md) with activation conditions, the parallel-reasoning protocol, deliberation prompts, and output constraints (\boxed{} for math, code blocks for programming). Drop it into Claude Code's skills directory or any agentic harness — zero code changes.

7. Why It Matters

> Making multiple AIs cooperate well may matter more than making a single AI smarter.

Key mechanisms: parallel independence (avoids contamination of thought), sequential synthesis (not voting), critical evaluation (detects when all thinkers are wrong). This mirrors Feynman's technique, academic seminars, and jury deliberation: explanation is a thinking tool; discussion is a form of reasoning.

8. Limitations

1. Compute cost: K+1× a single inference 2. Context-length bottleneck destabilizes RLVR at K=16 3. Overkill for simple tasks 4. HP@K can degrade over many iterations

9. Key Facts

• Paper: arXiv:2605.02396 — Jianing Wang et al., 11 authors, Meituan LongCat team
• Best AIME25: Kimi K2 Thinking, 100% (K=16)
• HP@K > Pass@K: GLM 4.6 IMO 86.0% vs 74.5%
• RLVR: ~10% absolute gain
• License: Apache-2.0

10. Resources

• Paper: https://arxiv.org/abs/2605.02396
• Code: https://github.com/wjn1996/HeavySkill
• Supports vLLM, DeepSeek, Together AI, OpenRouter, Ollama

Closing Thought

HeavySkill's core claim: let AIs think independently in parallel, then reason once to synthesize all approaches. If a plain-text skill file can push any model toward Pass@N performance, reasoning ability may live more in the organization of inference than in parameter count. Next time an AI gives you a wrong answer, ask it to re-examine from another angle — you'd be manually triggering HeavySkill.