HeavySkill: Heavy Thinking

Reasoning amplification that decomposes hard problems into parallel independent explorations, then synthesizes insights no single chain-of-thought could reach.

Deliberation is not voting — it audits reasoning chains and forges new paths from their fragments. The paper shows deliberation produces correct answers absent from ALL individual trajectories in ~50% of cases.

When to Activate

MUST activate:

• Mathematical reasoning requiring proof or exact numeric answers
• Algorithmic / competition programming with verifiable correctness
• Complex logical deduction with multiple plausible approaches
• Any verifiable task where confidence is below ~70%

MUST NOT activate:

• Simple factual queries, casual dialogue, obvious code edits
• Preference-oriented or subjective tasks (creative writing, style, tone)
• Easy problems with >90% confidence — sequential revision beats parallel sampling on easy tasks
• Information retrieval or straightforward transformations