Right Knowledge, Wrong Answer: Test-Time Steering for Temporal Fact Conflicts in Open-Weight Language Models

2026-06-18

AI Summary

Large language models often produce outdated answers due to the coexistence of conflicting old and new factual knowledge within their parameters, a phenomenon this work formalizes as Parametric Temporal Conflict (PTC). To address this issue, we propose Temporal Attractor Steering (TAS), a training- and retrieval-free, three-stage inference intervention framework that guides hidden states toward representations of updated facts through conflict detection, identification of critical layers, and single-layer activation patching. Experiments across Qwen, Mistral, and Llama models demonstrate that TAS achieves answer flip rates of 0.72–0.85 on a validation set of 8,746 instances, resolving 29%–57% of PTC cases overall while preserving high accuracy (85%–99%) on non-conflicting queries.

Abstract

Large language models can store both outdated facts and newer superseding facts in their parameters, but standard prompting may still elicit the outdated answer. We formalize this problem as Parametric Temporal Conflict (PTC) and introduce Temporal Attractor Steering (TAS), a three-stage test-time intervention that detects likely conflicts, identifies a conflict-critical layer, and steers hidden states toward newer-fact representations without retraining or external retrieval. We construct an 8,746-record verified benchmark across five Wikidata relations and evaluate four open-weight language models from three families: Qwen-2.5-1.5B/7B, Mistral-7B-v0.3, and Llama-3.1-8B. Single-layer activation patching achieves answer-flip rates of 0.72-0.85 across all models. End-to-end TAS resolves 29-57% of PTC cases while preserving 85-99% accuracy on non-conflict queries, outperforming a matched ITI baseline on three of four models. These results show that outdated parametric knowledge can be selectively overridden at inference time.

Problem

• Temporal Fact Conflicts
• Parametric Knowledge
• Outdated Information
• Language Models
• Test-Time Intervention

Innovation

• Temporal Attractor Steering
• Parametric Temporal Conflict
• test-time intervention
• activation patching
• fact updating