Research map / Architectures and efficiency / Attention and transformers

Taming Knowledge Conflicts in Language Models

ICML 2025 · Architectures and efficiency › Attention and transformers · approach: Attention head mechanisms · 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.

Identifies "superposition of contextual information and parametric memory" in attention heads and proposes JuICE, a training-free dual-run head intervention that steers LMs toward parametric or contextual knowledge.

Problem

Language models face knowledge conflicts between parametric memory and contextual input; prior memory-head/context-head views assume exclusive head roles, which fails across conflict types, and steering methods are limited to single conflict types.

Approach

Attention heads serve memory and context in superposition, so single-pass head knockout/scaling is unstable; JuICE selects heads with consistent effects across conflict types and, in a second run, adds scaled activations saved from a first run, yielding reliable steering without fine-tuning.

Key results

• On Gemma-2b, JuICE raises average parametric-belief accuracy under coherent conflict (type 3) from 10.5% (prompt) to 73.4% across six factual datasets.
• On contextual tasks with Gemma, JuICE averages 66.2% versus 45.0% original and 55.8% for the best PH3 variant.
• JuICE outperforms the single-run JuNe ablation on average, with about a 20-point gap for Gemma on parametric tasks.

Measured effects reported in the paper

Surprising findings

• Half of the top-ranked memory heads from substitution conflicts act as context heads under coherent conflicts.
• Combining individually effective head interventions in a single pass can counteract each other and reduce performance.

Limitations

• Evaluation focuses on factual recall and two contextual benchmarks; inter-context and intra-memory conflicts are out of scope.
• Hyperparameters (K, scaling factors) are tuned on validation data per model.

Experimental setup

Datasets: World Capital Athlete Sport Book Author Official Language Company Headquarter Company Founder NQ-Swap MemoTrap

Benchmarks: World Capital Athlete Sport Book Author Official Language Company Headquarter Company Founder NQ-Swap MemoTrap

Metrics: exact match accuracy

Baselines: Prompt PH3 CAD JuNe (single-run ablation)

Builds on: PH3 (memory and context heads) contrastive decoding (CAD) induction heads

Closest papers by meaning

• Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.91)
• ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
• Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (ICML 2025 · similarity 0.88)
• KScope: A Framework for Characterizing the Knowledge Status of Language Models (NeurIPS 2025 · similarity 0.87)
• The Atlas of In-Context Learning: How Attention Heads Shape In-Context Retrieval Augmentation (NeurIPS 2025 · similarity 0.87)
• $\texttt{ConflictBank}$: A Benchmark for Evaluating the Influence of Knowledge Conflicts in LLMs (NeurIPS 2024 · similarity 0.87)