{"id":"1a2ce464-f587-4365-975b-c949f0b68747","arxiv_id":"2505.07166","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The 'knowledge comes from pre-training' pattern seen in DPR/BERT does not generalize to Contriever (mean pooling) or RepLlama (decoder), where fine-tuning instead reduces neuron activation breadth.","lead":"This paper retests the claim that dense retrieval fine-tuning only reorganizes pre-trained knowledge, extending it from BERT/DPR to Contriever and Llama-based retrievers. It finds the original pattern holds for DPR but flips for mean-pooled and decoder-based models, so the earlier conclusion is architecture-dependent.","discovery_kind":"replication","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neuron-attribution analysis is underspecified: Eq. 4 uses an undefined scalar output P_x for vector-embedding models, and the Llama threshold is chosen post hoc, so the claim that decentralization is architecture-dependent is not yet supported.","rationale":"The reader's weakest assumption already identifies the neuron-attribution method as the fragile link, and I agree with that diagnosis. My stress-test sharpens it: the method is not merely \"possibly unreliable\" in the abstract; Equation (4) is underspecified for the models studied. A dense retriever maps text to a vector, so \"the model's scalar output P_x\" must be defined. The paper does not define it, and the independent processing of inputs in Section 5.2.1 rules out a natural relevance-score interpretation. The Llama threshold change (0.01 vs 0.1) is explicitly post hoc and can create the very reversal the paper reports. If the public code reveals an arbitrary P_x, or if the threshold sweep changes the pattern, the central claim about architecture-dependent knowledge decentralization collapses. The DPR linear-probing reproduction and the RepLlama deep-layer accuracy jump are useful and survive this concern, but the abstract's broader claim about neuron activation and non-universality is not currently supported. Because this is addressable by specification and validation, I would not reject the paper outright; the reader's CONDITIONAL verdict remains appropriate, with the condition strengthened to require a precise definition of P_x and a uniform threshold analysis. A secondary overstatement is that the abstract says \"pre-trained knowledge underpins retrieval performance\" without measuring retrieval effectiveness, but the attribution issue is the more load-bearing problem.","tokens_in":14650,"tokens_out":7748,"duration_ms":75491,"concrete_test":"Inspect the public repository (github.com/ielab/DenseRetriever-Knowledge-Acquisition) and identify the exact computation of P_x in Equation (4). Then rerun the neuron-attribution experiments for Figures 5–8 with (a) P_x defined as the L2 norm of the output embedding, (b) P_x defined as the cosine similarity between the query/passage embedding and a fixed reference embedding, and (c) a uniform threshold of 0.1 for all models plus a threshold sweep (e.g., 0.01, 0.05, 0.1, 0.2). If the DPR-vs-Contriever/Llama activation reversals in Figures 7–8 disappear or invert under any of these definitions, the non-universality claim in the abstract is an artifact of the chosen scalarization and threshold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central non-universality result rests on the integrated-gradient (IG) neuron-attribution analysis. In Section 5.1, Equation (4) defines each neuron attribution as the path integral of the gradient of \"the model's scalar output P_x\". A dense retriever's output is a vector embedding, not a scalar; Section 5.2.1 says questions and passages are processed independently, so P_x cannot be a query-passage relevance score. The paper never states what P_x is. If P_x is an arbitrary scalarization (e.g., one embedding coordinate or the norm), the attribution values—and hence the observed \"decentralization\" in DPR and its absence in Contriever/Llama—are not interpretable as knowledge organization. The threshold choice compounds this: all models use 0.1 × max except Llama, where 0.01 is adopted because it \"provided the clearest trend\" (Section 5.6). A post hoc threshold can manufacture the reported reversal. The authors themselves concede in Section 5.7 that neuron attribution \"may actually be not reliable for investigating dense retrieval.\" Since neither the DPR decentralization nor the Contriever/Llama reversal is supported without this method, the abstract's central claim outruns the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a reproducibility study of Reichman and Heck's claim that DPR fine-tuning does not add new knowledge but rather decentralizes knowledge that already exists from pre-training. The authors reproduce the original linear-probing and integrated-gradient (IG) neuron-activation analyses on DPR with a BERT backbone, and extend the experiments to the MS MARCO dataset, to mean-pooled Contriever, and to decoder-only RepLlama/Llama with EOS-token pooling. They report that the DPR query encoder reproduces the original trend, while the DPR passage encoder, Contriever, and RepLlama show different activation patterns, and they conclude that the original decentralization finding is architecture- and representation-dependent. The implementation and preprocessed datasets are publicly released.","tokens_in":14929,"tokens_out":4638,"duration_ms":43294,"significance":"If the IG-based conclusions are valid, the paper makes a useful contribution by showing that the decentralization effect is not universal across dense retrieval architectures, and it provides a reproducible extension of a previously single-model study. Concrete strengths include the public code and data, the addition of a second dataset, and the use of statistical significance testing for the linear-probing results. However, the central non-universality claim rests on a neuron-attribution method whose scalar output is left undefined for dense retrievers, the Llama activation threshold is chosen post hoc, the IG results have no quantified uncertainty, and the authors themselves concede in Section 5.7 that the method may be unreliable for dense retrieval. The paper's significance therefore depends on issues that must be resolved before the main claim can be accepted.","major_comments":[{"comment":"The quantity P_x in Eq. (4) is never defined, and for dense retrievers it is not obvious what it should be. The manuscript states in §5.2.1 that questions and passages are processed independently, and the model output is a dense vector embedding rather than a scalar relevance score. If P_x is an arbitrary scalarization (a single coordinate, a norm, or a projection), the resulting attributions and active-neuron percentages cannot be interpreted as knowledge organization without justification. Because Figures 5–8 are the sole evidence for the central claim that decentralization is architecture-dependent, this needs to be fixed by specifying P_x, providing a rationale for the scalarization, and ideally validating the IG attributions against an independent measure.","section":"§5.1, Eq. (4)"},{"comment":"The activation threshold is 0.1×max for all models except Llama-based models, where 0.01 is used because it \"provided the clearest trend.\" This is post hoc selection of the very quantity that defines the reported outcome, so the reversal for RepLlama in Figure 8 may be an artifact of threshold choice. The authors should report results for a common threshold across all models, quantify sensitivity over a range of thresholds, and provide uncertainty estimates, such as variability across examples or repeated runs.","section":"§5.6 and footnote 10"},{"comment":"The authors explicitly acknowledge in Section 5.7 that neuron attribution, the method on which their main generalization claim rests, \"may actually be not reliable for investigating dense retrieval.\" This admission applies equally to the DPR decentralization result and to the Contriever/Llama reversals. The linear probing results alone do not measure knowledge reorganization; they measure discriminative capacity. As written, the abstract's conclusion that fine-tuning \"primarily adjusts neuron activation rather than reorganizing knowledge\" and that this pattern is architecture-dependent outruns the evidence. The authors should either validate the method for dense retrieval or explicitly reframe the conclusions as conditional on the attribution method.","section":"§5.7 and Discussion"},{"comment":"The Discussion states that IG analysis shows DPR fine-tuning increases intermediate-layer activations by 32–41%, but no quantitative table or confidence interval supports this number in the main text, and the figures show aggregate percentages without error bars. This quantitative claim should be either backed by a table with variance estimates or removed.","section":"§7 Key Findings"}],"minor_comments":[{"comment":"There are numerous typos and naming inconsistencies: 'Contriver' appears after RQ3, 'ReplLlama' and 'RepLlama' are used interchangeably, 'LLAMA' appears in Figure 8, and 'bert-based-unacsed' appears in the captions of Figures 2 and 4.","section":"Throughout"},{"comment":"The reference list contains obvious placeholder citations ('Dai and Other 2022', 'Hao and Coauthors 2021', 'Smith and Doe 2021', 'Lee and Kumar 2021', 'Sun et al. 2020') and inconsistent entries; these must be corrected before publication.","section":"References"},{"comment":"The description of the epoch/validation procedure ('within a epoch of 50') is unclear; please specify how the best linear-probing layer is selected and applied.","section":"§4.2.3"},{"comment":"Figures 2, 3, and 4 state that a t-test 'between each DPR model with bert-based-unacsed backbone model' is marked with *; for the Contriever and Llama comparisons there is no DPR model, so the caption wording should be corrected.","section":"Figure captions"},{"comment":"RepLlama is described as 'castorini/repllama-v1-7b-lora-passage'; please clarify whether the comparison uses the LoRA adapter or the merged model and how the Llama backbone is run.","section":"§4.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable reproducibility contribution, but its main novelty—the architecture dependence of knowledge decentralization—is exactly where the evidence is weakest: the undefined P_x in Eq. (4), the post hoc Llama threshold, and the authors' own admission of the method's unreliability. I would ask the editor to weigh the fact that the authors themselves flag the central method's reliability, and I see no misconduct concern, only a need to tighten the claims and provide the missing validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real reproducibility contribution: it confirms the DPR trend Reichman and Heck reported, finds that Contriever and RepLlama do not show the same decentralization pattern in integrated-gradient attributions, and shows DPR-query and DPR-passage behave differently—an observation absent from the original work. Second, the central non-generalization claim is weaker than the abstract makes it sound, because the attribution method is never pinned down: Equation 4 defines P_x as a 'scalar output' but a dense retriever produces a vector embedding, and the text never says what scalarization was used. The post hoc threshold for Llama (0.01 vs 0.1) and the missing error bars compound the issue. The authors admit in Section 5.7 that neuron attribution 'may actually be not reliable for investigating dense retrieval.'\n\nWhat it does well: it ships code and data, tries to reconstruct the original setup, acknowledges reproduction difficulties (hard negative sourcing, linear probing details), and extends to a second dataset and two additional architectures. The linear probing results stand on their own: BERT-CLS retaining discriminative power, RepLlama's deeper-layer jump being a real and interesting architecture effect. The DPR reproduction alone makes this worth a serious referee.\n\nWhere the soft spots are: the IG analysis, which carries the decentralization reversal. The stress-test concern is on target. Since P_x is undefined and the threshold is tuned to show the desired trend, the comparison between DPR, Contriever, and Llama is not interpretable as evidence about knowledge organization. That does not kill the linear probing contribution, but it means the abstract's first sentence—'pre-trained knowledge underpins retrieval performance'—is a claim the paper never actually tests, because no retrieval effectiveness is measured. Also, the t-tests in figures are between model and backbone, not between conditions, so 'statistically significant' markers are easy to over-read.\n\nWho it is for: people working on dense retrieval interpretation and anyone treating Reichman and Heck as a general result. It deserves peer review, not desk rejection; a good referee would push for a defined scalar target, threshold sensitivity analysis with confidence intervals, and an abstract that matches what was measured. I would send it to review, conditionally. The core DPR finding is likely solid; the architecture-dependence claim needs either a defensible attribution target or softer phrasing.","headline":"A genuinely useful reproducibility extension that confirms and complicates Reichman and Heck, but the neuron-attribution analysis at the center of the non-generalization claim is too underspecified to carry it.","tokens_in":15472,"tokens_out":2280,"would_cite":true,"duration_ms":20606,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Pre-training, not fine-tuning, supplies DPR retrieval knowledge — but not for every retriever.","keywords":["dense retrieval","pre-training","fine-tuning","knowledge acquisition","knowledge decentralization","linear probing","neuron attribution","reproducibility study"],"falsifier":"Run the same neuron-activation comparison with an alternative attribution method, such as causally ablating individual intermediate-layer neurons and measuring retrieval effectiveness; if DPR-query does not show greater sensitivity to intermediate layers than the backbone, the decentralization pattern is an artifact of the integrated-gradient estimator. A simpler check: vary the activation threshold from 0.01 to 0.5 and see whether the reported 32–41% intermediate-layer increase persists.","tokens_in":14432,"feed_emoji":"🧠","tokens_out":5457,"duration_ms":48924,"temperature":0.7,"pith_summary":"This paper tries to establish whether the claim that dense-retrieval fine-tuning does not add knowledge—only reallocates it across neurons—holds beyond the single setup where it was first made. The authors reproduce the original experiments on BERT-based DPR, extend them to MS MARCO, and then test mean-pooled (Contriever) and decoder-only (RepLlama) retrievers. They find the original pattern reproduces for DPR query encoders: pre-trained BERT already separates relevant from irrelevant passages, and fine-tuning decentralizes neuron activations in intermediate layers. The pattern reverses for Contriever and RepLlama, where fine-tuning narrows the set of active neurons. The upshot is that the pre-training-versus-fine-tuning question has no architecture-independent answer.","feed_headline":"Fine-tuning reshapes, not adds, DPR knowledge","feed_subtitle":"Reproducibility study shows the pattern reverses for mean-pooled and decoder-only retrievers.","key_machinery":"Two analysis tools carry the argument. Linear probing trains a simple classifier on frozen layer-wise embeddings to test whether fine-tuning adds discriminative knowledge beyond pre-training. Integrated-gradient neuron attribution computes each feed-forward neuron's contribution to the final embedding (integrating weights from zero to their learned values, normalizing per example, and thresholding at 0.1 times the maximum, with 0.01 used for Llama models) to measure whether knowledge is centralized or decentralized before and after fine-tuning.","core_discovery":"The paper's central claim is that DPR fine-tuning does not introduce new retrieval knowledge into a BERT backbone; instead it changes how existing knowledge is accessed, primarily by increasing activations in intermediate feed-forward layers (32–41% more neurons above threshold) while suppressing early output-layer activations. This reproduces the earlier study's conclusion on both Natural Questions and MS MARCO for the DPR query encoder. But the same analysis on a mean-pooled retriever (Contriever) and a decoder-based retriever (RepLlama) shows the opposite: fine-tuning reduces the breadth of activated neurons. The paper therefore concludes that knowledge decentralization is architecture- and pooling-dependent, not a general property of dense retrieval, and that the passage encoder does not follow the query encoder's trend.","pith_inferences":["A testable extension: if knowledge decentralization is a real mechanism, causal interventions (ablating or amplifying intermediate-layer neurons) should change DPR retrieval effectiveness more than the same interventions on Contriever; the current evidence is correlational.","The authors' own caveat that neuron attribution may actually not be reliable for investigating dense retrieval suggests the principal robust finding may be methodological: linear-probing and attribution results are sensitive to pooling, backbone, and threshold choices, so claims about where knowledge lives need multi-method confirmation.","The RepLlama deeper-layer accuracy jump hints that EOS-token representations accumulate relevance signal late in the network; one could test whether layer-wise pruning of late layers degrades RepLlama more than DPR, which would make the architectural difference practically exploitable.","An implicit practical consequence: the cost of pre-training may be justified for BERT-style encoders, but decoder-based retrievers may benefit more from fine-tuning or representation design, so training budgets should depend on backbone family."],"forward_implications":["If DPR's query encoder relies on pre-trained knowledge, then for BERT-based retrievers, improving retrieval effectiveness beyond the backbone's ceiling requires pre-training changes, not just more contrastive fine-tuning.","If the decentralization effect reverses for mean-pooled and decoder-only retrievers, then research conclusions drawn from DPR alone cannot be used to guide training of Contriever-style or LLM-based retrievers.","Because the DPR passage encoder behaves differently from the query encoder, the original claim that fine-tuning affects both encoders the same way is weakened, and passage-side knowledge flow needs its own analysis.","Architecture and pooling strategy can matter more than fine-tuning for how discriminative the final embeddings are, as seen in RepLlama's 18–22% deeper-layer accuracy jump over its backbone."],"supporting_citations":[{"why":"Supplies the original hypothesis and the linear-probing and neuron-activation methodology that this paper reproduces and extends.","marker":"Reichman and Heck [2024]"},{"why":"Provides the DPR models and the official NQ hard-negative data used to build the linear-probing training sets.","marker":"Karpukhin et al. [2020]"},{"why":"Provides Contriever, the mean-pooled retriever used as an alternative fine-tuning recipe.","marker":"Izacard et al. [2022]"},{"why":"Provides RepLlama, the decoder-based retriever used to test backbone generalization.","marker":"Ma et al. [2024]"},{"why":"Foundational integrated-gradients method used to compute neuron attributions.","marker":"Sundararajan et al. [2017]"},{"why":"Supplies the knowledge-neuron attribution technique applied to feed-forward weights.","marker":"Dai and Other [2022]"},{"why":"Supports treating intermediate and output feed-forward layers as key-value knowledge storage, justifying the layer selection.","marker":"Geva et al. [2021]"}],"fun_headline_variants":["DPR fine-tuning adjusts neurons, not knowledge","Retrieval knowledge gain is architecture-dependent","Fine-tuning's role in dense retrieval varies by model","Pre-training knowledge is key for DPR, not all retrievers","Study: knowledge acquisition varies with encoder and pooling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The integrated-gradient neuron-attribution procedure, including the choice of activation thresholds, is a valid way to locate and compare knowledge distribution in dense retrievers; the authors themselves note this method may not be reliable for this setting.","fun_headline_variants_meta":{"raw":{"variants":["DPR fine-tuning adjusts neurons, not knowledge","Retrieval knowledge gain is architecture-dependent","Fine-tuning's role in dense retrieval varies by model","Pre-training knowledge is key for DPR, not all retrievers","Study: knowledge acquisition varies with encoder and pooling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3786,"prompt_tokens":958,"completion_tokens":2828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2753}},"tokens_in":574,"tokens_out":2828,"duration_ms":23056,"temperature":1.0,"reasoning_tokens":2753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:22:45.837038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same neuron-activation comparison with an alternative attribution method, such as causally ablating individual intermediate-layer neurons and measuring retrieval effectiveness; if DPR-query does not show greater sensitivity to intermediate layers than the backbone, the decentralization pattern is an artifact of the integrated-gradient estimator. A simpler check: vary the activation threshold from 0.01 to 0.5 and see whether the reported 32–41% intermediate-layer increase persists.","supporting_citations":[{"cited_title":"Dense passage retrieval for open-domain question answering","cited_arxiv_id":null,"evidence_quote":"Provides the DPR models and the official NQ hard-negative data used to build the linear-probing training sets."},{"cited_title":"Knowledge neurons in pre-trained transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the knowledge-neuron attribution technique applied to feed-forward weights."},{"cited_title":"Transformer feed-forward layers are key-value memories","cited_arxiv_id":null,"evidence_quote":"Supports treating intermediate and output feed-forward layers as key-value knowledge storage, justifying the layer selection."}],"review_version":1}