Verification Loops

Why verification beats retries.

The Problem

Single-pass generation assumes outputs are correct. Models often miss their own errors, and retrying produces different (not necessarily better) results.

Research

"Towards Autonomous Mathematics Research" (Google DeepMind, arXiv:2602.10177)

Finding: Generator-Verifier-Reviser pattern achieves 90% on IMO-ProofBench (math proofs) vs 67% base.

System:

Problem → Generator → Candidate → Verifier → [Pass | Revise | Restart]

Results:

• 90% on IMO-ProofBench Advanced (vs 67% base)
• Autonomous solutions to 4 open Erdős problems
• Natural language verifier admits failure

Why It Works:

1. Separate concerns — generator creates, verifier validates
2. Different perspectives — verifier approaches from critique angle
3. Explicit checking — verifier justifies pass/fail
4. Admits uncertainty — prevents false positives

arXiv

"Accelerating Scientific Research with Gemini" (Google, arXiv:2602.03837)

Finding: Human-AI collaboration with adversarial review detects subtle flaws.

Techniques:

• Iterative refinement
• Problem decomposition
• Adversarial reviewer to detect flaws
• Neuro-symbolic loop for code verification

arXiv

Quantitative Impact (IMO-ProofBench)

Approach	Accuracy
Base Model	67%
Retry (3x)	72%
Verification Loop	90%

Note: These results are from mathematical proof benchmarks. Your mileage may vary for code tasks.

Tachikoma's Implementation

Verifier-Code-Agent

1. GENERATE — Produce initial solution
2. VERIFY — Check with explicit criteria
3. REVISE — Fix based on verification
4. [Loop max 3 iterations]
5. REFLECT — Question approach, flag issues

Reflection-Orchestrator

1. Initial Output
2. Self-Critique
3. Revision
4. Final Verification
5. Reflect on confidence

Skill Chain

implement-verify:
  skills:
    - code-agent # Generate
    - verifier-code-agent # Verify
    - formatter # Clean up

When to Use

Use verification when:

• Complex implementations
• High-stakes fixes
• First-time features
• Correctness is paramount

Skip verification when:

• Simple tasks (<50 lines)
• Prototypes
• Well-understood patterns

See Also

• Skill Chains
• Research Overview