{"id":"59b114f2-d3b5-4b75-a730-9fe00d6a2dff","arxiv_id":"2411.11344","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On memorized question-answer pairs from KMIR and NQ, bottleneck and prefix adapters trained on entity-swapped contexts let a GPT-2 reader follow the new context most of the time, though no baselines are reported.","lead":"This short paper tests whether small trainable adapters can make a question-answering language model follow new context instead of repeating answers it memorized during training. It reports that two adapter types improve the model's switch to context-provided answers on two QA datasets, but without baselines or released code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The p_s metric cannot distinguish knowledge override from context-copying; original-answer accuracy and non-memorized held-out conflicts are never reported.","rationale":"The paper's idea is plausible and the qualitative examples in Tables I and II do show context-following behavior on a few cases. However, the quantitative evidence does not establish the central claim because the reported metric is essentially a measure of copying the substituted entity from the context on a self-selected memorized subset, with no baseline and no measurement of performance on original examples. The reader's weakest assumption was about external validity: whether the anti-factual substitution distribution matches real-world conflicts and whether flip rates transfer to broader faithfulness. My concern is more internal: even within the paper's own setup, p_s cannot distinguish knowledge override from a trivial context-copying strategy. This is a more fundamental flaw because it undermines the interpretation of the reported numbers, not just their generalizability. Adding original-answer accuracy and non-memorized held-out conflicts would directly test the mechanism. I found no evidence of formal verification, code release, or baselines that would independently support the claim. The existing REJECT verdict is consistent with this analysis, so I do not recommend changing it.","tokens_in":8890,"tokens_out":4129,"duration_ms":49863,"concrete_test":"Repeat the Section V.B evaluation with two added measurements: (1) exact-match accuracy on the original, unsubstituted versions of the same test questions before and after adapter training; (2) p_s on anti-factual examples built from questions that were NOT in the memorized subset, drawn from the remaining KMIR/NQ pools. If original-answer accuracy drops substantially, or if p_s on non-memorized conflicts is close to a copy-the-entity baseline, the claim that adapters override parametric knowledge fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing problem is in Section V.B and the conclusion in Section VI. The quantitative metric p_s is computed over entity-substituted examples constructed from the same memorized subset used to create adapter training data, and it only checks whether the model's output equals the substituted answer planted in the context. No results are reported for accuracy on the original, unsubstituted versions of those questions after adapter training, and no comparison is made against a simple baseline such as a context-copying heuristic or standard fine-tuning on the same anti-factual data. A high p_s is therefore equally consistent with the model having learned a shallow 'copy the entity mentioned in the context' behavior as with genuinely overriding parametric knowledge. The central claim that bottleneck adapters and prefix-tuning adapters 'effectively override memorized parameterized knowledge' requires evidence that the adapter changes the competition between parametric memory and context; the reported experiments only show that, on a specially selected set of 593/637 memorized items, the final model often outputs the entity placed in the context. Transfer to non-memorized questions and preservation of normal QA accuracy are untested, so the claim is not supported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to mitigate entity-based knowledge conflicts in QA by freezing a GPT-2 victim model and training bottleneck adapters or prefix-tuning adapters on 'anti-factual' triples, where the gold answer is replaced by a same-type entity and the context is rewritten correspondingly. It defines a memorization rate R_mem and a substituted-answer accuracy p_s, reports high p_s numbers on KMIR and NQ memorized subsets, and concludes in Section VI that these methods 'effectively override memorized parameterized knowledge.' The paper does not provide code, baselines, hyperparameters, or error bars.","tokens_in":9074,"tokens_out":4975,"duration_ms":49863,"significance":"If the central claim were established, the idea of using parameter-efficient adapters to steer a frozen model away from parametric memory and toward presented context would be a useful contribution to hallucination mitigation. The paper is clear about the problem formulation and the construction of anti-factual data, and it gives a concrete, falsifiable metric p_s. However, as it stands the evidence is not sufficient: the evaluation is confined to a specially selected memorized subset, the metric does not distinguish context-following from context-copying, and the promised 'minimum performance sacrifice' is never measured. The strengths are the clean experimental template and the parameter-efficient approach; the weaknesses are the missing controls and incomplete method specification.","major_comments":[{"comment":"The evaluation metric p_s only checks whether the model's output equals the substituted entity that is already present in the context. Because the adapter is trained to maximize exactly this conditional probability on anti-factual triples, a high p_s is equally consistent with a shallow 'copy the entity in the context' behavior as with genuinely overriding parametric knowledge. To support the Section VI claim, the paper must report at least: (i) accuracy on the original, unsubstituted questions after adapter training (to measure the 'minimum performance sacrifice' promised in Section I), (ii) p_s on non-memorized or held-out conflict examples, and (iii) a comparison against a context-copying baseline and standard full fine-tuning on the same anti-factual data. Without these, the central claim is not established.","section":"Section V.B, Eq. (p_s)"},{"comment":"The 'adversarial way' of training is never specified: there is no adversarial loss formula, no adversary architecture, no training schedule, and no hyperparameters. This is a load-bearing part of the method because it is what allegedly distinguishes the proposal from ordinary supervised fine-tuning, yet it is not reproducible. In addition, Section III explicitly says the paper reports 'our tentative schedule for the project,' which suggests the method section describes a proposal rather than a completed and validated approach.","section":"Section III.C"},{"comment":"The qualitative example in Table I is inconsistent with the quantitative metric. For the Heinz Hohner question, the context contains 'French Navy,' the bottleneck adapter output is 'France,' and the text says the model 'successfully output the shuffled answer.' Under the exact-match definition of p_s in Section V.B, 'France' would not count as the substituted answer. The paper needs to reconcile this discrepancy: either the metric underreports the method's behavior (e.g., fine-grained or semantic matching should be used) or the qualitative claim is inaccurate.","section":"Section V.A, Table I"},{"comment":"The test set is constructed by taking the 593/637 memorized samples from the training split and substituting the answer, so p_s is evaluated on the same distribution used to create the adapter training data. No evidence is given that improvements on this memorized subset transfer to broader QA faithfulness or to naturally occurring knowledge conflicts, which is the stated goal in the abstract. The paper needs an independent prediction—for example, behavior on non-memorized questions or on conflicts sampled from a realistic retrieval distribution—before the hallucination-reduction claim can be taken as supported.","section":"Section IV.B and Section V.B"}],"minor_comments":[{"comment":"The column headers 'accuracy on training set' and 'accuracy on test set' are misleading; these are p_s values on the anti-factual memorized subset, not general QA accuracy. Please rename them to make the scope explicit.","section":"Tables III and IV"},{"comment":"There are numerous typos and formatting issues, including 'Datset' in Section IV, 'extend' for 'extent' in Section V.B, 'bottle neck adapter' in Section III.C, and inconsistent spelling of 'parameterized/parameterised.'","section":"Throughout"},{"comment":"Several references appear unrelated to the claims they are attached to, for example [13] is cited to support a statement about hallucination risks but the reference is an alpha-mining model paper, and [17],[26],[33] are about resource allocation, transformer heart-disease prediction, and graph storage. The citation list should be cleaned so that each reference supports the specific statement.","section":"References"},{"comment":"The text refers to Figure 1 and contains a caption-like block 'FIG 1 THREE STAGES...', but no actual figure image appears in the manuscript. Either include the figure or remove the reference.","section":"Figure 1"},{"comment":"The notation is inconsistent: R_mem uses x'' for the predicted answer while p_s uses x-hat, and the indicator function notation is described loosely. Please unify the notation and define all variables at first use.","section":"Section III.B and Section V.B"}],"recommendation":"reject","confidential_remarks":"The manuscript reads as an early-stage project report rather than a completed study: the Methods section explicitly contains a 'tentative schedule,' no code or data are released, and the experimental section has no baselines or error bars. The reference list also contains several citations that appear unrelated to the asserted claims, which warrants a check on citation practices. I would not encourage resubmission in the current form; the central claim would need a substantially expanded evaluation with proper controls and full method specification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, this is a promising sketch, not a finished paper. The core idea—train a bottleneck adapter or prefix-tuning on anti-factual QA examples so the model follows context instead of its memorized parametric answer—is sensible and modestly new as a combination. The qualitative examples look consistent, and the framing in Longpre's entity-substitution framework is clean. But the quantitative support is too thin for the central claim in Section VI. The p_s metric only checks whether the model outputs the substituted answer; because the same anti-factual distribution was used to build the training data and the test set, a high p_s is equally consistent with the model learning to copy the entity in context, not with genuinely overriding parametric knowledge. Original-answer accuracy after adapter training is never reported, so we don't know whether normal QA degrades. There are no baselines: no standard fine-tuning on the same anti-factual data, no context-copying heuristic, no comparison to K-adapter style injection. The adversarial loss mentioned in Section III.C is never specified. Error bars, hyperparameters, and code are all absent. On the positive side, the idea is plausible and the evaluation is at least on pre-selected memorized samples, which is the right testbed for the intended behavior. The sample sizes (593 and 637) are small but not fatal if the effect were large and clean. The biggest soft spot is the inability to rule out shallow copying. I don't think this deserves a full-length review yet. It could be a workshop paper or the seed of a real study. For peer review, I would send it back for major revision: report accuracy on original unmodified questions, add a standard fine-tuning baseline, specify the adversarial loss, and test on non-memorized conflicts to show transfer. As written, the evidence does not support the 'overriding memorized parameterized knowledge' claim. Who gets value from this? Someone working on parameter-efficient knowledge editing might find the combination worth trying, but not as a citable result.","headline":"A plausible adapter-based idea for overriding memorized QA answers that is undermined by an evaluation unable to distinguish context-copying from knowledge override.","tokens_in":9586,"tokens_out":2736,"would_cite":false,"duration_ms":26315,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training small bottleneck adapters or prefix-tuning adapters on entity-substituted question-answer triples makes a QA model override its memorized parametric answers and follow the supplied context instead, reducing a common form of…","keywords":["knowledge conflicts","question answering","hallucination mitigation","adapter tuning","prefix tuning","parametric knowledge","anti-factual substitution","frozen backbone"],"falsifier":"Run the adapter-trained model on naturally occurring conflicting passages, such as an NQ passage stating a fact newer than or different from the answer the original model memorized, without any engineered entity substitution, and measure exact-match faithfulness; if $p_s$ stays high on engineered swaps while natural conflicts still receive the memorized answer, the central claim fails.","tokens_in":8706,"feed_emoji":"🧠","tokens_out":7522,"duration_ms":65242,"temperature":0.7,"pith_summary":"The paper targets a failure mode in open-book question answering: a reader model that has memorized an answer can ignore retrieved context and repeat the memorized entity, producing a hallucination when the context says otherwise. The authors construct anti-factual triples by swapping the gold answer for another entity of the same type, then train either a bottleneck adapter or a prefix-tuning adapter on these triples while keeping the base GPT-2 frozen. They report that both adapters make the model switch to the substituted answer on the memorized subset, reaching about 92 percent exact-match accuracy on KMIR and about 64 percent on Natural Questions. The paper's intended conclusion is that parameter-efficient adapter training can steer a model toward context-grounded output without full fine-tuning.","feed_headline":"Two small adapters override a QA model's memorized answers","feed_subtitle":"Bottleneck and prefix adapters trained on swapped entities lift context-following accuracy to about 92 percent on KMIR.","key_machinery":"The load-bearing mechanism is the combination of an anti-factual dataset and a frozen-backbone adapter: a fine-tuned GPT-2 victim model is scored by a memorization rate that counts exact matches of original answers, a subset of memorized questions is rebuilt with same-type entity substitutions, and then cross-entropy training is run on the substituted triples while only the adapter parameters (bottleneck adapter or prefix-tuning vectors) are updated. The evaluation metric $p_s$ measures how often the model emits the substituted answer, which the paper takes as evidence that the context rather than parametric memory is driving generation.","core_discovery":"The central discovery is that knowledge conflicts can be mitigated at the adapter level: with the base language model frozen, a small trainable module trained on substituted entity contexts is enough to override the model's memorized answer. The method builds an anti-factual dataset of triples $(q, c', x')$ where $x'$ is a same-type substitution of the original answer $x$ and $c'$ is the context with $x$ replaced by $x'$, then maximizes the language-modeling likelihood of $x'$ conditioned on $(q, c')$ through the adapter parameters only. On examples the victim model previously answered correctly, the fraction of substituted answers generated after adapter training reaches 92.9 percent on KMIR and 64.1 percent on Natural Questions with the bottleneck adapter, and 91.9 percent and 64.8 percent with prefix tuning. The authors conclude that both methods can effectively override memorized parameterized knowledge.","pith_inferences":["The paper does not establish that flip rate on engineered anti-factual substitutions transfers to natural hallucination; a testable extension would measure faithfulness on unmodified or naturally conflicting contexts, not only swapped entities.","Because adapters are small and additive, one could train separate context-trusting adapters for different knowledge domains and switch them at inference time, an application the paper does not explore.","The accuracy gap between KMIR (about 92 percent) and NQ (about 64 percent) suggests context complexity matters; a natural extension is to test whether longer or more adversarial contexts require larger adapters or additional training signal."],"forward_implications":["If the result holds, a small adapter trained on anti-factual examples can replace full fine-tuning for reducing entity-level hallucination, preserving the base weights and lowering compute.","The high substituted-answer accuracy indicates that models can be steered to trust retrieved context even when it contradicts a memorized fact, which is directly relevant to retriever-reader question-answering pipelines.","Because both bottleneck and prefix adapters achieve similar accuracy, the effect is not tied to one adapter architecture, suggesting a general property of small trainable modules.","The method targets cases where the model has a crisp memorized answer, so improving these cases should reduce clear-cut factual hallucinations of the exact-match type."],"supporting_citations":[{"why":"Supplies the entity-based knowledge-conflict framework and the corpus-substitution setting used to build the anti-factual dataset.","marker":"[38]"},{"why":"Provides the adapter-based knowledge-injection formulation that the bottleneck adapter approach builds on.","marker":"[30]"},{"why":"Provides the KMIR benchmark and triple-completion data used for the first memorization and adapter experiments.","marker":"[37]"},{"why":"Provides the Natural Questions dataset used for the second set of memorization and anti-factual experiments.","marker":"[39]"},{"why":"Defines the bottleneck adapter layer used as one of the two knowledge-erasing modules.","marker":"[31]"},{"why":"Inspires the adversarial-style training of the adapters for more robust performance.","marker":"[32]"}],"fun_headline_variants":["Adapters beat memorized QA answers with swapped facts","Small adapters override QA memory via anti-factual data","Frozen LM plus adapters counter knowledge conflicts","Anti-factual adapters fix QA hallucination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation rests on the assumption that the anti-factual, same-type substituted dataset represents real-world knowledge conflicts and that raising substituted-answer accuracy on a pre-selected memorized subset reduces hallucination in general question answering.","fun_headline_variants_meta":{"raw":{"variants":["Adapters beat memorized QA answers with swapped facts","Small adapters override QA memory via anti-factual data","Frozen LM plus adapters counter knowledge conflicts","Anti-factual adapters fix QA hallucination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1268,"prompt_tokens":859,"completion_tokens":409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":347}},"tokens_in":475,"tokens_out":409,"duration_ms":4378,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:37:26.794559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the adapter-trained model on naturally occurring conflicting passages, such as an NQ passage stating a fact newer than or different from the answer the original model memorized, without any engineered entity substitution, and measure exact-match faithfulness; if $p_s$ stays high on engineered swaps while natural conflicts still receive the memorized answer, the central claim fails.","supporting_citations":[{"cited_title":"Entity-based knowledge conflicts in question answering,","cited_arxiv_id":null,"evidence_quote":"Supplies the entity-based knowledge-conflict framework and the corpus-substitution setting used to build the anti-factual dataset."},{"cited_title":"K -adapter: Infusing knowledge into pre -trained models with adapters,","cited_arxiv_id":null,"evidence_quote":"Provides the adapter-based knowledge-injection formulation that the bottleneck adapter approach builds on."},{"cited_title":"KMIR: A Benchmark for Evaluating Knowledge Memorization, Identification and Reasoning Abilities of Language Models","cited_arxiv_id":"2202.13529","evidence_quote":"Provides the KMIR benchmark and triple-completion data used for the first memorization and adapter experiments."},{"cited_title":"Natural questions: a benchmark for question answering research,","cited_arxiv_id":null,"evidence_quote":"Provides the Natural Questions dataset used for the second set of memorization and anti-factual experiments."},{"cited_title":"Efficient transfer learning for quality estimation with bottleneck adapter layer,","cited_arxiv_id":null,"evidence_quote":"Defines the bottleneck adapter layer used as one of the two knowledge-erasing modules."},{"cited_title":"Improving text -image matching with adversarial learning and circle loss for multi -modal steganography,","cited_arxiv_id":null,"evidence_quote":"Inspires the adversarial-style training of the adapters for more robust performance."}],"review_version":1}