Research map / Large language models / Retrieval and knowledge

Can LLMs Take Retrieved Information with a Grain of Salt?

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 0 citations (OpenAlex)

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.

Defines context-certainty obedience as the total-variation closeness of an LLM's output to an ideal mixture of its prior and context answer, evaluates eight open-weight LLMs on ClashEval, and proposes a training-free interaction strategy (prior reminder, certainty recalibration, context simplification) that cuts obedience error by about 25%.

Problem

LLMs are not known to adapt their answers appropriately to the expressed certainty of retrieved context, which matters in high-stakes retrieval-augmented QA. The paper asks whether models obey external certainty scores.

Approach

Formalize the ideal response as a mixture (1-c)·prior + c·context-answer distribution, measure deviation via integrated total variation distance over certainty levels, and improve obedience purely through prompt/interaction design: reminding the model of its own prior answer, recalibrating the certainty values shown in the prompt, and simplifying the context to the extracted answer.

Key results

• Full interaction strategy reduces average context-certainty obedience error from 0.52 to 0.39 (about 25%) across eight LLMs without weight updates.
• Gemma-3-27B error drops from 0.48 to 0.28 and Llama-3.3-70B from 0.48 to 0.29 with the full strategy.
• Prior reminders alone reduce average error from 0.52 to 0.46; recalibration and simplification gains are roughly additive.

Surprising findings

• Models almost always fail to recall their prior responses after seeing a context, even when told to ignore 0%-certain contexts.
• Elaborated or longer contexts increase believability, so models overtrust complex contexts even at low certainty.

Limitations

• Requires output logits, so it cannot be applied to API-based models or reasoning models.
• Evaluation restricted to short-answer settings where the context conveys a definite answer and filtered to samples retrieved by all models, dropping Gemma-1B samples substantially.
• Only quantized large models (INT4 AWQ) were tested on a single A100.

Experimental setup

Datasets: ClashEval

Benchmarks: ClashEval

Metrics: context-certainty obedience error (epsilon_obey, integrated per-certainty total variation distance

Baselines: No-intervention baseline prompting (certainty score in prompt) Prior reminder only Ground-truth reminder variant

Builds on: ClashEval AWQ

Closest papers by meaning

• ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence (NeurIPS 2024 · similarity 0.89)
• Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.89)
• To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.88)
• Information Discernment in Large Language Models (NeurIPS 2026 · similarity 0.87)
• Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.86)
• FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows" (ICLR 2025 · similarity 0.86)