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
Execution
Stage 1: Parallel Reasoning
Spawn K independent agents in a single message (parallel). Each solves from scratch with zero sibling knowledge.
| K | When |
|---|---|
| 3 | Standard — most problems |
| 5 | High-stakes — competition math, critical correctness |
Agent prompt:
Solve this problem step by step. Show your complete reasoning and arrive at a final answer. Use whatever approach you find most natural — algebraic, geometric, constructive, brute force, or proof by contradiction.
Problem: {query}
Requirements:
- Reason from first principles, show all work
- Final answer clearly marked
- Math: \boxed{answer} | Code: code block
Stage 2: Memory Cache
Before deliberating, structure the outputs:
- Collect all K trajectory outputs
- Shuffle trajectory order (prevents position bias)
- Prune if exceeding token budget — truncate reasoning, preserve final answers
Stage 3: Sequential Deliberation
You perform this yourself — never delegate deliberation.
- Classify query type (math / code / logic / multi-step)
- Map answer distribution — what answers appear, how often?
- Audit each chain — logic valid? gaps? sign errors? hidden assumptions?
- Cross-validate — do independent approaches confirm the same result?
- Apply critical skepticism:
- Majority is signal, not proof
- Minority answer with tighter logic may be correct
- All may be wrong — re-derive from fragments if needed
- Synthesize the definitive answer
Stage 4: Output
Final answer only — no meta-analysis unless asked:
- Math:
\boxed{answer}| Code: code block | General: prose - Match the user's query language
Iterative Deliberation
Trigger when first deliberation produces low confidence or fundamental disagreement.
- Concatenate deliberation as additional trajectory:
cache(t) = deliberation(t-1) ∥ trajectories - Re-run Stage 3 on augmented cache
- Maximum 2–3 iterations — performance degrades beyond that
Anti-Patterns
| Pattern | Fix |
|---|---|
| Voting instead of deliberating | Audit reasoning chains, not answer counts |
| Delegating deliberation to agent | Always deliberate yourself — full context needed |
| Sequential agent spawning | Spawn all K in one parallel message |
| Skipping shuffle | Always randomize order before deliberating |
| Over-iterating (>3 rounds) | Stop at confidence or iteration 3 |
| Activating on subjective tasks | Gate: is there a verifiable answer? |
| Using on easy problems | If >90% confident, just solve directly |
References
Load when deeper understanding needed:
references/landscape.md— How HeavySkill compares to Best-of-N, Self-Consistency, Forest-of-Thoughtreferences/compute.md— Cost analysis, K selection, non-monotonic performancereferences/tensions.md— Key design tensions: consensus vs minority, width vs depth, speed vs thoroughnessreferences/paper-details.md— Paper methodology details: memory cache strategies, trajectory selection, RLVR