- The paper synthesizes three research programs showing that attribution methods provide different explanations of predictions, fact dynamicity predicts resistance to contextual updates better than popularity, and real-world RAG behavior differs from synthetic benchmarks.
- LLMs often resist retrieved information about frequently changing facts but readily overwrite stable facts, highlighting the need for better measures of memory conflict and more targeted context-integration strategies.
- Real-world evidence from DRUID shows that direct, authoritative, post-claim evidence improves context use, while long contexts and weak query-context similarity reduce faithfulness; no single failure signal or context-manipulation method works universally.
Overview
This paper presents a keynote delivered at ECIR 2025 by Isabelle Augenstein (University of Copenhagen), surveying her group's research on how LLMs combine parametric knowledge — facts encoded in weights during pre-training — with contextual knowledge supplied at inference time, as in retrieval-augmented generation (RAG). The keynote is structured around three research threads: attribution of parametric knowledge underlying predictions, diagnostic probing of knowledge conflicts, and characterization of when RAG systems actually use retrieved context. A recurring theme is that the interplay between these two knowledge sources remains underexplored, particularly the question of when contextual knowledge should overwrite parametric knowledge.
Motivation: factuality limits of LLMs
The introduction grounds the problem in well-documented LLM shortcomings: dataset contamination means benchmark performance often reflects answer recall rather than reasoning, with drastic drops under small paraphrases; low-resource languages and most reasoning types remain weak; and factual errors stem from the absence of an external knowledge base. Existing remedies each carry costs. Chain-of-thought prompting and self-consistency checking are undermined by prompt instability; continual learning is expensive; knowledge editing suffers from ripple effects beyond the intended edit, removal of long-tail knowledge due to knowledge superposition, and the nontrivial task of identifying what to edit. Inference-time correction, modularized knowledge-grounded frameworks, and RAG address context-dependence but presuppose that models will actually integrate retrieved content — which studies show they frequently do not, especially when context conflicts with pre-trained memory. This motivates the three lines of work summarized below.
Attributing parametric knowledge
The first study [2402.xxxxx via yu-etal-2024-revealing] asks what parametric knowledge influences a prediction, comparing two previously isolated families of attribution methods — instance attribution (identifying influential training examples) and neuron attribution (identifying important neurons) — within a unified evaluation framework. The framework aligns influential training instances with important neurons and assesses faithfulness through sufficiency tests (activating key neurons), completeness tests (suppressing them), and fine-tuning on influential instances. Analysis focuses on MLP classification layers for tractability, justified by prior evidence of high correlation with overall model weights.
Three findings stand out. First, most MLP neurons can be removed without significantly changing predictions, contradicting the common assumption that most model knowledge resides in these neurons; the authors hypothesize attention weights play a larger role in knowledge encoding. Second, fine-tuning on instances identified as highly influential performs no better than fine-tuning on an equal number of randomly selected instances — a regression test attributes this to the significantly lower diversity of "influential" instances. Third, the combined use of instance and neuron attribution is more effective than either alone at discovering dataset artifacts, specifically overfitting to lexical patterns. Notably, the two attribution families yield different explanations of the same prediction, so conclusions about a model's inner workings depend heavily on method choice.
Probing intra-memory and context-memory conflicts
Because attribution methods are expensive, require full parameter access, and can produce contradictory findings on partial parameters, the second line of work uses probing tests instead. The DynamicQA dataset [marjanovic-etal-2024-dynamicqa] distinguishes static facts (single representation), temporal facts (changing over time), and disputable facts (viewpoint-dependent). Dynamic facts induce intra-memory conflict — uncertainty from contradicting representations in the training data — while conflicting context under RAG induces context-memory conflict. The former is measured with semantic entropy; the latter with a novel "coherent persuasion score" approximating the semantic shift in output distribution given competing context.
The central result is counter-intuitive: manipulated contexts for static and low-dynamicity facts persuade LMs more than dynamic facts. Facts that change regularly are less likely to be updated via context retrieval, while immutable facts are easily overwritten. Fact dynamicity emerges as the strongest, most consistent negative indicator of model persuasion across models, outperforming fact popularity, which prior RAG work had used as a guiding signal. This directly challenges popularity-based heuristics and indicates that new measures of intra-memory conflict and better indicators of successful context usage are needed, especially in low-certainty domains.
Characterizing realistic context usage in RAG
The third thread addresses the gap between synthetic context-usage studies and real-world RAG conditions. The DRUID dataset [hagstrom-etal-2025-reality] provides claim verification with real-world contexts, paired with a novel context usage measure (ACU). Several patterns emerge: fact-check source context yields high usage, plausibly due to assertive, direct language and multi-argument discussion; evidence published after the claim and gold evidence sources increase usage; references to external sources correlate weakly; LLMs prioritize high query-context similarity, which is harder to obtain in real retrieval settings; and faithfulness degrades for long contexts.
A key reality check follows from comparing DRUID to synthetic datasets: synthetic data oversells the impact of certain characteristics such as knowledge conflicts, which are rare in realistically retrieved data, and exaggerates context repulsion. Moreover, no single characteristic reliably indicates RAG failure in real-world settings. The implication is that findings from synthetic benchmarks may not transfer, and real-world-aligned evaluation is necessary.
Follow-up work extends this in two directions. A multistep knowledge interaction analysis [islam2025multistepknowledgeinteractionanalysis] finds task-dependent behavior: in claim verification, where conflicts are frequent, models prefer contextual over parametric knowledge; in commonsense QA with fewer conflicts, they lean on parametric memory. A benchmarking study of context manipulation techniques (CMTs) [hagström2025cub] — spanning prompting through mechanistic interventions — finds no clear winner among techniques, though larger LLMs use context better on average, and smaller models can match or exceed them when paired with the CMT best suited to that specific model.
Limitations and open questions
The keynote is candid that its results raise more questions than they resolve. The neuron-attribution analysis was restricted to MLP classification layers for computational reasons, leaving open whether conclusions hold across full model weights. The finding that influential-instance fine-tuning matches random selection rests on a diversity-based regression explanation rather than a mechanistic account. The dynamicity result establishes correlation with persuasion but does not yet yield actionable RAG design rules, and the authors explicitly call for new measures of intra-memory conflict. The DRUID analysis identifies multiple failure indicators but no single predictive one, limiting diagnostic utility. Finally, the CMT benchmarking shows no dominant intervention strategy, meaning practitioners currently lack principled guidance for resolving context-memory conflicts.
Conclusion
The keynote consolidates three complementary research programs into a coherent account of the parametric–contextual knowledge interplay: attribution methods give divergent explanations of the same prediction; fact dynamicity, not popularity, best predicts resistance to contextual persuasion; and synthetic benchmarks systematically exaggerate phenomena that are rare in real retrieval settings. The author frames progress on these questions as requiring collaboration between information retrieval and NLP communities, noting recurring parallels between current LLM behaviors and those observed for earlier pre-trained LLMs — recitation over reasoning, and preference for easily digestible sources — suggesting continuity rather than discontinuity in model behavior across generations of architectures.
