Confirmation Lock in LLM Agents: Benchmark Reveals Cognitive Bias
Single LLM agents in a prompt-tool-analysis-response loop show systematic bias, much like confirmation bias in psychology. The model generates a hypothesis, crafts tool queries to fit it, interprets results, and confirms its conclusions while ignoring contradictions. This is Confirmation Lock: the generator and critic merge in one context, blocking objectivity.
The experimental LOCK-R benchmark mathematically captures deviation from Bayesian ideals. Agents get a budget of 4 tool calls, a false initial anchor, and must output JSON with hypothesis probabilities (H1, H2, H3). A Bayesian oracle recalculates true probabilities based on evidence likelihoods. The key metric is Bayes Regret (R_mean): the difference in probability distributions.
A Python algorithmic bot achieves zero regret, validating the setup.
LOCK-R Benchmark: Rules and Metrics
The environment is fully deterministic:
- Three mutually exclusive hypotheses.
- Budget: 4 calls (search, verification).
- False anchor on H1.
- JSON with confidence % after each step.
Key metrics:
- Tool_loop_repetition_rate: Query repetition (up to 68% in locked agents).
- K_c (asymmetry coefficient): Response to confirmations vs. refutations. Ideal is 0; LLMs show strong negatives (stubbornness).
Tests on Qwen3.5-9B and GPT-5.4 revealed query collapse: agents loop on one tool, reinforcing bias.
Forced oracle_query_control (correct sources) boosts accuracy from 25% to 85%.
Key Insight: Asymmetric Evidence Processing
LLM agents react asymmetrically. Refutations drop confidence by 5–10%, while confirmations spike it to 90%. K_c goes negative: dismissing criticism, hyping support.
This isn't a context shortage—it's a structural flaw in single agents.
The Chain-of-Thought Paradox
Comparing Thinking (CoT) and Non-Thinking modes:
- Exploration (search/planning): CoT is essential; regret without it hits 2.26.
- Verification (weighing facts): CoT worsens it, increasing regret. The model burns tokens on lawyerly arguments, rationalizing contradictions.
CoT amplifies illusions over objectivity.
Solution: Blind Judge
Asymmetric pipeline: Thinking Explorer (search) + Non-Thinking Judge (verification).
The blind judge gets only raw facts (hypotheses + tool logs), no drafts. Regret drops from 1.47 to 0.09.
Real-World CodeTriageEnv Scenario:
- 500 error on payment page.
- False anchor: Redis CPU spike.
- True cause: JSONDecodeError from payment gateway.
| Mode | Accuracy | Unique Tools | Regret |
|-----------------|----------|--------------|--------|
| Single-agent | 40% | 2.8 | 0.25 |
| Blind checker | 100% | 2.0 | 0.07 |
Single agents chase Redis shadows, ignoring stack traces. The blind judge zeros in on JSON.
Frontier Models and Scaling
GPT-5.4 has lower baseline regret (0.03), but CoT doubles it to 0.06. The architectural flaw persists: scaling boosts guessing, not judging.
Key Takeaways
- Single agents are vulnerable to Confirmation Lock due to role fusion.
- CoT helps generation but harms verification.
- Splitting into Explorer + Blind Judge cuts regret to 0.09.
- Even top models need asymmetric pipelines.
- K_c metric quantifies LLM stubbornness.
— Editorial Team
No comments yet.