Research map / Large language models / Retrieval and knowledge
Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
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.
Builds a framework that elicits LLM parametric memory and generates coherent counter-memory to study how LLMs handle knowledge conflicts between external evidence and internal beliefs, finding high receptiveness to single counter-evidence but strong confirmation bias when mixed evidence is presented.
Problem
How receptive LLMs are to external evidence, especially counter-memory that conflicts with their parametric memory, which matters for tool- and retrieval-augmented LLMs.
Approach
Instead of heuristically editing parametric memory (entity substitution, negation), instruct an LLM to generate coherent counter-memory from scratch, with entailment and answer-consistency filters to ensure elicited memory is firm, then run controlled single- and multi-source conflict experiments.
Key results
- Single-source generated counter-memory makes LLMs highly receptive; entity-substitution counter-memory is mostly rejected by ChatGPT, GPT-4 and PaLM2.
- With both parametric and counter-memory present, GPT-4 reaches about 80% memorization ratio on the most popular PopQA questions, showing confirmation bias.
- Evidence order changes outcomes: PaLM2 and Llama2-7B show over 30% fluctuation in memorization ratio depending on which evidence is first.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide coherent counter-memory as the sole external evidence | LLMs · PopQA and StrategyQA | adoption of counter-memory | Helped |
| replace generated coherent counter-memory with entity-substitution counter-memory | LLMs · PopQA | adoption of counter-memory | Hurt |
| add parametric-memory evidence alongside counter-memory | LLMs · PopQA and StrategyQA | memorization ratio: 3.7 -> 99.8; 2.6 -> 99.6 | Helped |
| increase the proportion of parametric-memory-aligned evidence | LLMs · PopQA | memorization ratio: 3.7 -> 99.8 | Helped |
| put the parametric memory first instead of the counter-memory | ChatGPT · PopQA | memorization ratio: 46.7; 40.1 | Mixed |
| put the counter-memory first instead of the parametric memory | PaLM2 and Llama2-7B · PopQA and StrategyQA | memorization ratio: variations ... surpassed 30% | Mixed |
| increase the popularity of the queried entity | GPT-4 · PopQA | memorization ratio: 80% | Helped |
| add irrelevant evidence without relevant evidence | Llama2-7B · PopQA | irrelevant answer rate: 92.4 | Hurt |
| increase the amount of irrelevant evidence alongside relevant evidence | Llama2-7B · PopQA | irrelevant answer rate: 2.6 -> 13.1 | Hurt |
| replace counter-memory with the short counter-answer | ChatGPT · PopQA | counter-answer rate: 56.7% to 18.8% | Hurt |
| replace parametric memory with the short memory answer | ChatGPT · PopQA | memory answer rate: 42.7% to 43.9% | Helped |
| fragment the evidence instead of presenting it as a consolidated passage | ChatGPT · StrategyQA | preference shift to the other intact evidence: 38.2% | Hurt |
Surprising findings
- LLMs are highly receptive to coherent counter-memory even when it contradicts their parametric memory, contrary to prior work.
- Short counter-answer evidence is rarely adopted (18.8%), while short parametric-memory answers are adopted at a similar rate to full evidence, indicating asymmetric receptiveness.
Limitations
- Relies on ChatGPT for counter-memory generation and DeBERTa NLI for entailment checking, which may introduce bias.
- Evaluated only on PopQA and StrategyQA with multiple-choice reformulation, limiting the generality of findings.
Experimental setup
Datasets: PopQA StrategyQA
Benchmarks: PopQA StrategyQA
Metrics: memorization ratio answer distribution accuracy answer inconsistency rate
Baselines: entity substitution counter-memory
Builds on: entity substitution counter-memory memorization ratio PopQA StrategyQA
Closest papers by meaning
- $\texttt{ConflictBank}$: A Benchmark for Evaluating the Influence of Knowledge Conflicts in LLMs (NeurIPS 2024 · similarity 0.91)
- Taming Knowledge Conflicts in Language Models (ICML 2025 · similarity 0.91)
- KScope: A Framework for Characterizing the Knowledge Status of Language Models (NeurIPS 2025 · similarity 0.89)
- ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.89)
- From Static to Dynamic: Knowledge Metabolism for Large Language Models (AAAI 2024 · similarity 0.88)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.88)
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.