- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes ReNA, a detector for context–memory conflicts in RAG. Given a context, a question, and a short model-generated answer, it predicts whether the answer follows the context or reverts to the model’s parametric knowledge. For each training prompt, the authors teacher-force two reference answers under the same input: one supported by the context and one supported by parametric memory. They then compute positive gradient-times-activation attribution for every FFN neuron and every attention “neuron” (a value-coordinate within a head), normalize the scores into within-family shares, and take the difference between the two references. Neurons whose preference is directionally consistent across training examples and whose average contrast is large form a fixed per-layer “atlas.” This step uses only the references, not the actual responses or labels. At detection time, ReNA computes the same attribution for the actual response, reads out its shares on the atlas coordinates, and feeds them to a logistic-regression classifier. On the QA, MR, and MC subsets of ConFiQA and on a constructed SQuAD2 conflict set, with LLaMA-2-7B-Chat and LLaMA-2-13B-Chat, ReNA reaches average AUROCs of 0.936 and 0.927, above SEP and FrequencyAware. Causal experiments show that scaling the selected neurons shifts the contextual-versus-parametric logit margin more than matched random neurons, but that patching their activations restores the original preference less well than random neurons do.
Critical strengths
-
The paper defines its units of analysis and attribution procedure precisely, and it separates atlas discovery (references only) from classifier fitting (actual responses and labels) cleanly.
-
The same-prompt reference contrast is a sensible design. It controls for input-side variation when comparing two answer trajectories.
-
The causal analysis is careful and honest. The authors use matched random controls, paired bootstrap intervals, and multiple-comparison correction, and they report the negative sufficiency result instead of hiding it.
-
The matched-random-atlas ablation is a good control, and the authors discuss its implications openly.
Critical weaknesses
-
The task may be close to trivial, and the paper skips the obvious simple baselines. The prompt limits answers to five words, so the actual response almost always equals one of the two references. The SQuAD2 construction makes this explicit: a parametric example must satisfy r_act = A_p exactly, and a faithful example must match the gold answer. In ConFiQA, the context-supported answer appears verbatim in the context, while the parametric answer usually does not. A baseline that just checks whether the response appears in (or overlaps with) the context could plausibly reach near-perfect AUROC. I’d also want to see a linear probe on hidden states at the same positions, and the difference in response log-likelihood with versus without context. Among the current baselines, only SEP is supervised, so the comparison mostly pits a supervised classifier against unsupervised methods.
-
The hyperparameters change across settings without a stated selection protocol, and the results are sensitive to them. On 7B, the authors use K = 64/32 with γ = 0.7 on ConFiQA, K = 256/128 with γ = 0.55 on SQuAD2, and change the regularization C to 0.03 for that one setting only. On 13B, they use K = 512/256 with γ = 0. The paper never explains how it picked these values. If the authors chose them by looking at cross-validation scores without nested selection, the reported numbers are optimistic. Table A.1 also has inconsistencies: two 7B rows carry the † marker; 128/64 beats the chosen 64/32 on ConFiQA AUROC (0.9714 vs. 0.9638); and the text says 13B ConFiQA Macro-F1 peaks at 256/128, while the table shows 512/256 is higher (0.9144 vs. 0.8982). On 7B SQuAD2, AUROC swings between 0.73 and 0.85 across atlas sizes.
-
The evidence that specific neuron identities matter is weak, and the causal analysis studies a different neuron set from the detector. A random atlas with matched layer and family composition already reaches 0.888 AUROC on ConFiQA, so most of the signal seems to come from attribution shares in general rather than from the selected identities. The intervention experiments don’t use the detector’s atlas. They use direction-balanced sets ranked by a Wilson lower bound, with 64 attention and 128 FFN neurons per layer (a 1:2 ratio), whereas the detector uses 64/32 (2:1). The scaling effect grows steadily with budget up to 6,144 neurons per layer, and patching sufficiency falls below random on all four datasets. I read these results as saying the selected neurons aren’t especially privileged, which sits awkwardly with the paper’s “atlas” framing.
-
The scope and practicality are limited. The paper only tests LLaMA-2 models, single-entity short answers, and one type of conflict, far from realistic RAG settings with long responses and partial hallucinations. Each detection needs a backward pass over the response, and the paper reports no compute cost. Building the atlas requires a known parametric answer for every training example. The cross-domain transfer test only freezes the atlas and retrains the classifier on the target domain, and the three ConFiQA subsets share the same construction, so the test says little about generalization.
Questions for authors
- How do a simple “response appears in context” or lexical-overlap baseline, and a linear probe on hidden states, perform on these datasets?
- Does the 13B model use γ = 0? If so, why do ReNA and Magnitude Only differ in Table 2 for 13B?
- How did you choose K, γ, and C for each model–dataset setting? Was the selection nested inside the cross-validation folds?