Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge

TL;DR

TRACK introduces a two-stage evaluation framework (knowledge probing and knowledge injection) to study how LLMs propagate updated facts through multi-step reasoning when conflicts arise with internal parametric knowledge. It instantiates three realistic scenarios (WIKI, CODE, MATH) and proposes novel metrics—Full Knowledge Entailment (FKE) and Holistic Pass (HP)—to separately assess faithfulness of reasoning and final answer correctness. Across a diverse set of models, findings show that providing updated facts can yield limited or even negative gains, with performance often degrading as more conflicting facts are supplied, due to both integration failures and flawed downstream reasoning. The benchmark provides a rigorous, real-world evaluation tool to guide future development of knowledge-aware reasoning in LLMs, including considerations for continual updates and multi-domain reasoning.

Abstract

A common solution for mitigating outdated or incorrect information in Large Language Models (LLMs) is to provide updated facts in-context or through knowledge editing. However, these methods introduce knowledge conflicts when the knowledge update fails to overwrite the model's parametric knowledge, which propagate to faulty reasoning. Current benchmarks for this problem, however, largely focus only on single knowledge updates and fact recall without evaluating how these updates affect downstream reasoning. In this work, we introduce TRACK (Testing Reasoning Amid Conflicting Knowledge), a new benchmark for studying how LLMs propagate new knowledge through multi-step reasoning when it conflicts with the model's initial parametric knowledge. Spanning three reasoning-intensive scenarios (WIKI, CODE, and MATH), TRACK introduces multiple, realistic conflicts to mirror real-world complexity. Our results on TRACK reveal that providing updated facts to models for reasoning can worsen performance compared to providing no updated facts to a model, and that this performance degradation exacerbates as more updated facts are provided. We show this failure stems from both inability to faithfully integrate updated facts, but also flawed reasoning even when knowledge is integrated. TRACK provides a rigorous new benchmark to measure and guide future progress on propagating conflicting knowledge in multi-step reasoning.

Figures (5)

• Figure 1: An illustration of Track's two-stage evaluation framework using a multi-hop QA example. (i) Knowledge Probing: We first decompose a complex question into atomic facts to identify a model's specific knowledge gaps. (ii) Knowledge Injection: We then provide correct facts as conflicting knowledge to test LLM's ability to propagate it within its reasoning.
• Figure 2: Full illustration of the Track benchmark across our three reasoning scenarios: Multi-Hop QA (WIKI), Code Generation (CODE), and Mathematical Reasoning (MATH). The benchmark follows a two-stage process. (i) Knowledge Probing: we identify the model's knowledge gaps by testing it on the required atomic facts, the individual pieces of knowledge needed to solve the complex question. (ii) Knowledge Injection: we evaluate reasoning by comparing a closed-book setting (using only the model's internal knowledge) with an open-book setting (where the identified knowledge gaps are provided). Performance is assessed using our metrics: Answer Pass (AP), Full Knowledge Entailment (FKE), and Holistic Pass (HP).
• Figure 3: Knowledge Confidence (KConf) distributions of Known and Unknown facts across (\ref{['fig:kconf_by_scenario']}) reasoning scenarios, and (\ref{['fig:kconf_by_model']}) model series. A consistent color scheme is used across both panels (Blue: Known facts, Red: Unknown facts).
• Figure 4: Failure analysis at the atomic fact level for the Append method, averaged across all models. The flow shows the percentage of all facts that were initially known or unknown (left) and their outcome in the reasoning (right): successfully integrated (Entail), being the first point of failure (Direct Failure), or failing due to a previous error (Error Propagation).
• Figure 5: The data generation pipeline for our three Track scenarios. Each pipeline sources from Wikidata vrandevcic2014wikidata, BigCodeBench zhuo2024bigcodebench, or PRM800K lightman2023lets, and produces five key components: the multi-step reasoning question $q$, final answer $a$, atomic facts $K_q$, probing questions $q_i$, and probing answers $a_i$. The generation method for each component is denoted by an icon: for deterministic scripts, for LLM-based generation, and both for a hybrid approach. Components with no icon are directly adapted from existing benchmarks.