Research map / Large language models / Retrieval and knowledge
Conformal Information Pursuit for Interactively Guiding Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Conformal Information Pursuit (C-IP), which selects interactive queries by minimizing the average conformal prediction set size (linked to conditional entropy) instead of unreliable LLM-probability entropy, and evaluates it on 20 Questions and MediQ with open-weight instruction-tuned LLMs.
Problem
Sequential query selection for LLM agents in interactive question answering is hampered by miscalibrated LLM output probabilities, which make conditional entropy estimates for information-gain-based query selection inaccurate and yield suboptimal query chains and predictions.
Approach
Use the average size of split-conformal prediction sets as a distribution-free uncertainty proxy: an entropy upper bound (via a proposition relating prediction set size to conditional entropy) is minimized at each iteration to pick the next query, with coverage marginalized over query-answer chain length.
Key results
- On 20Q with binary closed queries, C-IP achieves higher accuracy than IP and random selection at each iteration (target coverage 0.9).
- On 20Q with free-text queries, C-IP (alpha=0.1) reaches 0.83 success rate vs 0.65 for IP and 0.57 for UoT (binary), with average length 10.58.
- On MediQ, C-IP performs comparably to IP across Internal Medicine, Pediatrics and Neurology, and is competitive with non-sequential prediction while offering interpretable query chains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace IP entropy-based query selection with C-IP prediction-set-size selection | Llama-3.1-8B · 20 Questions, open queries, free-text answers | success rate: 0.65 $\pm$ 0.18 -> 0.83 $\pm$ 0.05 | Helped |
| Replace IP entropy-based query selection with C-IP prediction-set-size selection | Llama-3.1-8B · 20 Questions, closed queries, free-text answers | average query length: 13.81 $\pm$ 0.16 -> 8.87 $\pm$ 0.31 | Helped |
| Replace IP entropy-based query selection with C-IP prediction-set-size selection | Llama-3.1-8B · 20 Questions, closed queries, binary answers | success rate: 0.29 $\pm$ 0.11 -> 0.29 $\pm$ 0.04 | No clear effect |
| Replace IP entropy-based query selection with C-IP prediction-set-size selection | Llama-3.1-8B · 20 Questions, open queries, binary answers | success rate: 0.31 $\pm$ 0.04 -> 0.37 $\pm$ 0.11 | Helped |
| Use UoT lookahead instead of C-IP immediate information-gain selection | Llama-3.1-8B · 20 Questions, open queries, binary answers | success rate: 0.37 $\pm$ 0.11 -> 0.57 $\pm$ 0.13 | Helped |
| Replace binary answers with free-text answers | C-IP · 20 Questions, open queries | success rate: 0.37 $\pm$ 0.11 -> 0.83 $\pm$ 0.05 | Helped |
| Vary desired coverage α (ablation) | C-IP · 20 Questions and MediQ | predictive performance | No clear effect |
| Vary estimation sample count n_est across {1, 2, 4, 8} (ablation) | C-IP with Llama-3.1-8B · 20 Questions, closed binary and open binary queries | accuracy | No clear effect |
| Vary the number of sampled open queries per iteration (ablation) | C-IP with Llama-3.1-8B · 20 Questions, open queries | predictive performance | No clear effect |
Surprising findings
- Under binary answers, UoT achieves higher success rate (0.57) than C-IP (0.37) on open queries, indicating that look-ahead beats immediate-reward selection in that setting.
- C-IP undercovers on Neurology MediQ and on some open-query settings, and in Qwen2.5 and Phi-3 open-query experiments it did not outperform direct prompting.
Limitations
- Requires sampling to estimate prediction sets and entropy, and only a single-pass tractable approximation is left to future work.
- Coverage guarantee is marginalized over query-answer chain length rather than the actual history, so guarantees are not rigorous for the realized chains; undercoverage occurs in several settings.
- Relies on LLM instruction-following and accessible output logits.
Experimental setup
Datasets: Animals with Attributes 2 (AwA2) MediQ MedQA
Benchmarks: 20 Questions (AwA2 animals) MediQ
Metrics: accuracy per iteration average query length empirical coverage success rate average prediction set size
Baselines: Random query selection IP (entropy-based) Direct Prompting Uncertainty of Thought (UoT) Platt Scaling Temperature Scaling Non-sequential prediction Enumerate
Builds on: Information Pursuit Split Conformal Prediction V-IP Uncertainty of Thought
Closest papers by meaning
- Large language model validity via enhanced conformal prediction methods (NeurIPS 2024 · similarity 0.89)
- Prune 'n Predict: Optimizing LLM Decision-making with Conformal Prediction (ICML 2025 · similarity 0.88)
- Conformal Prediction Beyond the Seen: A Missing Mass Perspective for Uncertainty Quantification in Generative Models (NeurIPS 2025 · similarity 0.88)
- Conformal Language Modeling (ICLR 2024 · similarity 0.87)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.87)
- From Drift to Coherence: Stabilizing Beliefs in LLMs (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.