Research map / Large language models / Retrieval and knowledge
LACIE: Listener-Aware Finetuning for Calibration in 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 LACIE, a listener-aware DPO finetuning method that trains LLM speakers to express calibrated confidence (explicit and implicit) by optimizing against a simulated listener's acceptance decisions.
Problem
LLMs are overconfident when conveying answers, both through explicit confidence markers and implicit cues (tone, detail, backstories), which misleads human listeners into accepting wrong answers.
Approach
Cast calibration as preference optimization with a two-agent speaker-listener game: sampled speaker responses are judged by a listener model for acceptance and compared against gold correctness, with a preference function that rewards true accepts/rejects and penalizes false accepts, so the speaker learns to induce a calibrated listener distribution.
Key results
- Across Mistral-7B, Llama3-8B and Llama3-70B base models, LACIE improves induced listener AUROC by 20.7 points on average and reduces ECE by 7.8 points, with about 18% absolute precision gain.
- In a human study with 79/78 non-known items, LACIE reduced false accepts by 47% (32 to 17) while keeping true accepts roughly constant (31 to 30).
- On TruthfulQA (out-of-distribution, trained on TriviaQA), Mistral-7B truthfulness rises from 0.27 to 0.55 at a cost of about 9 points informativeness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply listener-aware preference finetuning | Mistral-7B base · TriviaQA | AUROC: 0.54\interval{0.00} -> 0.74\interval{0.02} | Helped |
| Apply listener-aware preference finetuning | Mistral-7B chat · TriviaQA | ECE: 0.23\interval{0.01} -> 0.09\interval{0.01} | Helped |
| Apply listener-aware preference finetuning | Llama3-8B base · TriviaQA | Precision: 0.55\interval{0.02} -> 0.70\interval{0.01} | Helped |
| Apply listener-aware preference finetuning | Llama3-70B base · TriviaQA | AUROC: 0.53\interval{0.02} -> 0.80\interval{0.02} | Helped |
| Apply listener-aware preference finetuning | Mistral-7B base · TriviaQA human evaluation | False Accept: 32 -> 17 | Helped |
| Apply listener-aware preference finetuning on TriviaQA | Mistral-7B · TruthfulQA evaluation split | Truthfulness: 0.27 -> 0.55 | Helped |
| Apply listener-aware preference finetuning on TriviaQA | Mistral-7B · TruthfulQA evaluation split | Informativeness: 0.99 -> 0.90 | Hurt |
| Use correctness-only preference finetuning instead of listener-aware preference finetuning | Mistral-7B base · TriviaQA | AUROC: 0.58 -> 0.74 | Hurt |
| Increase the number of training questions | LACIE with Mistral-7B listener · TriviaQA | Precision | Helped |
| Increase the number of training questions toward 10000 | LACIE with Mistral-7B listener · TriviaQA | Precision | Helped |
Surprising findings
- Abstention emerges from training even though no training examples contain abstentions.
- Speaker accuracy on all data drops after training mainly due to increased abstention, while accuracy on non-abstained examples stays comparable.
Limitations
- Preference function is conservative and may not reflect true user utility in downstream human-AI teams.
- Trained models can produce annoying or unpolite confident outputs emphasizing how obvious questions are.
- Evaluation is limited to trivia/TruthfulQA and a small human study (100 questions, 5 annotators).
Experimental setup
Datasets: TriviaQA TruthfulQA
Benchmarks: TriviaQA TruthfulQA
Metrics: AUROC ECE precision recall abstention rate truthfulness informativeness
Baselines: base model truthful-only finetuning chat/instruct models
Builds on: DPO QLoRA Rational Speech Acts
Closest papers by meaning
- Aligning Large Language Models for Faithful Integrity Against Opposing Argument (AAAI 2025 · similarity 0.89)
- SteerConf: Steering LLMs for Confidence Elicitation (NeurIPS 2025 · similarity 0.88)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.88)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning (ICML 2024 · similarity 0.87)
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.87)
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.