{"id":"de4521c5-2706-48b0-8d5b-40997686aa94","arxiv_id":"2608.10627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A probe trained only on classical context-memory conflict data detects decomposition-induced contradictions (DI-CC) in an LLM's fact-splitting step, while self-consistency detection fails.","lead":"This paper shows that when a language model is asked to split a passage into atomic facts before fact-checking, it can silently swap in its own beliefs and produce a claim that contradicts the passage it was supposed to summarize. This suggests the decomposition step of fact-checking pipelines is not always neutral, and can itself be a source of error.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-shot probe transfer may separate DI-CC from controls via claim position/length artifacts rather than a shared conflict mechanism, because NQ-Swap training uses a fixed prompt-token position while decomposition activations come from variable claim positions; no position/length matching is…","rationale":"The paper is unusually transparent: it reports the Mistral null, the sparse natural rate, the CAD side effects, the post-hoc variance-ratio selection, and the class-imbalance recall collapse. These disclosures are genuine strengths and should be credited. My concern is not about honesty or internal consistency; it is about whether the central mechanistic inference is secure.\n\nThe central inference is H0: a probe trained only on NQ-Swap conflict, at a fixed position in a QA prompt, separates DI-CC positions in decomposition output. For this to be evidence of shared mechanism, the probe must be responding to conflict-related content, not to incidental properties of where or how DI-CC claims appear. The manuscript does not report controlling for extraction position, claim length, or claim order. Residual-stream activations encode positional and surface information; a linear probe can exploit this. The appendix itself documents explicit-contrast phrasing and variable claim formats, so the confound is live. The Mistral failure makes it more, not less, important to rule out a Qwen-specific artifact, because that is exactly what a family-specific positional bias would look like.\n\nThe proposed test is decisive and cheap: fixed-position extraction and matched negatives. If it passes, the mechanistic claim is substantially strengthened. If it fails, the paper should be read as a behavioral characterization of DI-CC in Qwen2.5, not a proof of mechanism identity. I therefore keep the reader's CONDITIONAL verdict, with an added concrete condition.","tokens_in":17394,"tokens_out":6083,"duration_ms":62522,"concrete_test":"Rerun the H0 evaluation with activations extracted at a fixed token offset within each atomic claim (e.g., first token of the claim, or last token) rather than any claim-dependent position, and with DI-CC negatives matched to positives on claim length, claim index, and entity. Additionally rerun excluding the explicit-contrast DI-CC claims ('not 1949' phrasing) noted in Appendix 1.13.7. If AUC falls to near chance, the transfer signal was a positional/lexical artifact; if AUC remains around 0.86, the concern is rejected.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The H0 result is the only evidence for the central mechanistic claim, and it rests on a frozen probe trained on NQ-Swap activations extracted at the final prompt token before answer generation. For decomposition, the manuscript does not state the extraction position for each atomic claim, nor does it report controlling for claim length, claim index, or token offset. If DI-CC claims are systematically longer, later in the decomposition, or phrased with explicit contrast tokens (Appendix 1.13.7 reports six 14B cases phrased as 'born in 1912, not 1949'), the probe could separate on these correlates rather than on a conflict mechanism. The Mistral below-chance result shows the signal is not universal, but it does not rule out a Qwen2.5-specific position artifact; in fact, a family-specific positional bias would produce exactly such a pattern. Because H0 is the load-bearing bridge from 'behaviorally looks like conflict' to 'same mechanism as classical conflict,' an uncontrolled position/length confound would collapse the paper's central claim to a behavioral observation within one model family.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Decomposition-Induced Context-Memory Conflict (DI-CC), a failure mode in decompose-then-verify pipelines where a decomposer overrides source-text content with its own parametric knowledge, producing atomic claims that contradict the source. The central claim is mechanistic: DI-CC is the same underlying phenomenon as classical context-memory conflict, just shifted one pipeline stage earlier. The primary evidence is H0, a zero-shot linear probe transfer experiment: a probe trained only on NQ-Swap conflict activations separates DI-CC positions from faithful decompositions (AUC 0.881) and from non-contradictory elaborations (AUC 0.863) in Qwen2.5-7B, at n=64, with human verification of positive cases. The paper also reports that SelfCheckGPT-style self-consistency fails to detect DI-CC (AUC 0.508), that context-aware decoding suppresses DI-CC but with severe parsing side effects, and that the signal has a scale floor (3B null, 7B and 14B positive) and fails to replicate in Mistral-7B (AUC 0.397, below chance). The paper is unusually candid about its limitations, including post-hoc layer-selection, low natural occurrence rates, and the absence of replication on naturally occurring hallucinated text.","tokens_in":17494,"tokens_out":4041,"duration_ms":40371,"significance":"If the H0 result is correct, the paper makes a valuable contribution: it identifies decomposition as a non-neutral stage and ties a decomposition-specific failure to a well-studied phenomenon, with practical implications for fact-checking and long-form factuality evaluation. The strengths are real: code and data are released, positive cases are human-verified, multiple layer-selection criteria are reported, and the paper explicitly reports nulls and boundary conditions. However, the load-bearing mechanistic claim rests on a zero-shot probe transfer whose interpretation depends on controlling for position and surface-form confounds, and whose generality is contradicted by the Mistral result. As written, the evidence supports a Qwen2.5-family-specific behavioral and representational correlation more strongly than a general mechanistic identity claim. The paper's honest scoping partially mitigates this, but the central H0 statement is still broader than what the data establish.","major_comments":[{"comment":"The probe is trained on NQ-Swap activations extracted at the final prompt token before answer generation, but the manuscript does not state where activations are extracted for each atomic claim during decomposition. If DI-CC claims occur at systematically different token offsets, positions within the decomposition, or lengths than legit and DI-UE claims, the reported AUCs could reflect these positional or surface-form correlates rather than a conflict mechanism. This concern is concretized by Appendix 1.13.7, which reports six DI-CC cases phrased as explicit contrasts (\"born in 1912, not 1949\"), and by the fact that the comparison set is imbalanced (64 DI-CC vs. 1626 legit). The authors should report the extraction position, claim length, and claim index distributions for all three classes, and rerun the probe separation within strata matched on these variables, or otherwise show that the transfer is not driven by position or length.","section":"§4.1–4.2"},{"comment":"The Mistral-7B result (AUC 0.397, below chance, with a narrow CI excluding 0.5 in the wrong direction) and the Falcon-7B low-yield result show that the zero-shot probe transfer does not generalize across model families. The paper interprets this as a family-specific boundary, but this interpretation conflicts with the unqualified formulation of H0 in Section 1.3, which states the probe \"should\" separate DI-CC positions if DI-CC is mechanistically homologous to classical conflict. Since the classical context-memory conflict literature is not family-specific, a below-chance result in a second family either undermines the homology claim or requires a mechanistic account of why the same phenomenon would produce an inverted representation in Mistral. The authors should either restate H0 and the conclusion as explicitly Qwen2.5-family-specific, or provide such an account. As written, the central claim is stronger than the cross-family evidence allows.","section":"§8 (Cross-cutting)"},{"comment":"The definition of DI-CC already includes three criteria: the claim is not entailed by the source, the injected content is recoverable from the model's parametric knowledge, and the injection directly contradicts the source. This is nearly a restatement of context-memory conflict, so the behavioral labeling procedure builds the phenomenon into the data. The H0 probe transfer is then presented as independent mechanistic evidence, but it is subject to the concern that the probe is separating examples on the basis of features that the labeling pipeline used to define the classes (e.g., recoverability or contradiction-related surface patterns) rather than on a shared internal mechanism. The paper should clarify what additional, non-circular evidence the probe transfer provides beyond the definition, and ideally show that the probe's decision boundary is not reconstructible from the NLI labels or from simple lexical features of the claims.","section":"§2 (Problem Definition)"}],"minor_comments":[{"comment":"The abstract contains corrupted or truncated text: \"too sparss not manifest on naturally-occurring hallucinatedtext\" and \"detecablish\" and \"chhan overstate\"; these should be corrected.","section":"Abstract"},{"comment":"The text refers to \"Table 4.2b below\" when discussing the four (sample, layer-criterion) configurations, but the table is not actually present in the manuscript; the appendix table with the four rows is not labeled 4.2b. The cross-reference should be fixed.","section":"§4.2 (Table reference)"},{"comment":"The citation-verification passage mentions using \"the user's own research_gap_analysis.md literature-review notes\" to target searches; this is not a reproducible or publicly available resource and should be removed or replaced with a description of how the citations were independently verified.","section":"§1.12 (Citation Verification)"},{"comment":"The 14B AUC 0.983 is presented as the strongest result in the study, even though the variance-ratio criterion was adopted after seeing the accuracy-based null at 14B. The paper acknowledges this, but the figure and the narrative still give the post-hoc number visual prominence; it would be clearer to present the 14B variance-ratio result as explicitly exploratory, separate from the pre-registered accuracy-based evidence.","section":"§7 (Scale Ablation)"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent, and the within-family H0 result is likely real. However, the central mechanistic claim needs a position/length control analysis and a clearer scoping to the Qwen2.5 family before it can be accepted as stated. The Mistral below-chance result is more damaging than the paper frames it; it should either be explained mechanistically or the H0 claim should be narrowed. The manuscript also contains some odd presentation artifacts (e.g., the reference to a user-provided research-gap file) that suggest the text was assembled from multiple sources and needs editorial cleanup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know first: this paper is worth reading. It defines DI-CC, a failure mode where a decomposer overrides the source text with its own parametric belief, and provides evidence that this shares a mechanism with classical context-memory conflict, via a linear probe trained only on NQ-Swap data. The main result holds within Qwen2.5 under both layer-selection criteria, with human verification, and the paper is unusually honest about its limitations, including the Mistral failure and the post-hoc layer criterion. The negative results for SelfCheckGPT and the low natural rate are reported, not swept under.\n\nWhat is new: the formal definition and the zero-shot probe transfer test. The probe separating DI-CC from both faithful and elaborative claims without ever seeing decomposition data is a genuinely clever idea.\n\nSoft spots: the biggest one is the position/length confound. NQ-Swap training activations come from a fixed prompt token; the decomposition activations come from unspecified claim positions. If DI-CC claims are longer, later, or phrased with explicit contrast tokens (Appendix 1.13.7 shows six such cases), the probe could be picking up on those correlates. The paper doesn't report controls for claim length or index. That's a real gap, and a reviewer should push on it. It doesn't necessarily sink the paper—the separation from DI-UE argues against a generic 'anything unusual' detector—but it's unresolved.\n\nThe definition of DI-CC is also close to a restatement of context-memory conflict; that's not a flaw per se, but it means the conceptual novelty is thinner than the mechanistic test. The central result is confined to one model family; Mistral's below-chance result is a serious boundary, and the paper reads it correctly as an open question rather than a failure.\n\nThe abstract has some corrupted text ('sparss', 'chhan'), which should be fixed, but that's cosmetic.\n\nWho this is for: anyone building decompose-then-verify pipelines or studying knowledge conflict. It deserves a serious referee; the confound needs addressing, but the honest reporting and the novel test outweigh the gaps. I'd engage.","headline":"A transparent and mechanistically interesting paper on a real failure mode in decompose-then-verify pipelines, with a strong within-family probe result and a known but under-addressed potential confound.","tokens_in":18115,"tokens_out":3399,"would_cite":true,"duration_ms":31101,"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":"The paper claims that when a decompose-then-verify pipeline splits a passage into atomic claims, the decomposer can substitute its own parametric belief for what the source says, producing a claim that contradicts the source, and that…","keywords":["decomposition-induced context-memory conflict","context-memory conflict","hallucination detection","fact-checking pipelines","mechanistic interpretability","probing","context-aware decoding","self-consistency"],"falsifier":"A control probe trained on NQ-Swap with conflict labels randomly permuted should not separate DI-CC positions from faithful decompositions; if a shuffled-label probe reaches AUC near 0.86 on the same Qwen2.5 activations, the claimed mechanism-specific transfer is not established. Alternatively, the paper's own Mistral-7B result, below-chance AUC under the same protocol, provides a case where the predicted transfer fails, so any complete account must explain why a genuine mechanism produces above-chance transfer in one model family and reliably wrong-direction transfer in another.","tokens_in":17048,"feed_emoji":"🧩","tokens_out":6313,"duration_ms":54834,"temperature":0.7,"pith_summary":"This paper argues that the decomposition step in decompose-then-verify pipelines is not a neutral preprocessing stage. A decomposer language model can be induced to override what a source passage actually says with its own stored beliefs, producing an atomic claim that contradicts the passage it was supposed to summarize, a failure the paper names Decomposition-Induced Context-Memory Conflict (DI-CC). The paper's central claim is mechanistic: DI-CC is the same underlying phenomenon as classical context-memory conflict, just moved one pipeline stage earlier. To support this, a linear probe trained only on classical conflict data, never exposed to decomposition outputs, separates DI-CC positions from faithful decompositions and from non-contradictory elaborations at AUC 0.86-0.88. The paper also shows that self-consistency-based detection misses DI-CC, that a classical mitigation only partly transfers with serious side effects, and that the mechanism's natural scope is narrow.","feed_headline":"A fact-checking step can contradict the text it verifies","feed_subtitle":"When a decomposer splits a passage, its own beliefs can override the source; a cross-task probe spots it.","key_machinery":"The central object is a hypothesized linear direction $w$ in the residual stream, a representational context-memory-conflict signature, at some layer $\\ell^*$ that separates activations where the model's parametric belief conflicted with context from activations where it did not. The paper trains a per-layer logistic-regression probe on NQ-Swap activation vectors extracted at the final prompt token, freezes it, and applies it zero-shot to decomposition activations. The other load-bearing component is the operational definition of parametric injection, which uses an NLI model to establish non-entailment from the source and a closed-book knowledge dump to establish recoverability from the decomposer's own knowledge.","core_discovery":"On its own terms, the paper establishes a formally defined failure mode: a claim $c_i$ produced by decomposer $D$ from source $R$ shows parametric injection when content $I$ is not entailed by $R$ yet is stably recoverable from $D$'s closed-book knowledge. If $I$ directly contradicts $R$, it is DI-CC; if it merely fills a gap, it is DI-UE. The load-bearing test is H0: a probe $w$ trained only on classical context-memory conflict data should, applied zero-shot to decomposition activations, separate DI-CC generation positions from faithful claims and from DI-UE positions. The paper reports that it does: AUC 0.881 for DI-CC versus faithful and 0.863 for DI-CC versus DI-UE under variance-ratio layer selection at n=64, with permutation p<0.0005 and every positive case human-verified. The same probe signal does not appear at 3B scale, appears at 7B and 14B under the adopted criterion, and fails below chance on Mistral-7B, so the paper frames the mechanism as established within the Qwen2.5 family with an open cross-family question.","pith_inferences":["The probe-transfer result implies DI-CC should be detectable in any decompose-then-verify or claim-extraction task without retraining on that task's outputs; testing on RAG-style claim extraction from knowledge-intensive passages would be a direct extension.","If the mechanism is genuine, decomposition prompts that explicitly license fact-checking against parametric knowledge are the wrong default for faithful summarization, since they invite exactly the override the paper elicited.","The identity-fabrication side effect seen under CAD and in baseline coreference failures suggests a distinct vulnerability: when coreference lacks a strong anchor, the decomposer can substitute an unrelated identity, a failure mode that may require separate detection.","The self-consistency failure shows that stability across resamples should not be equated with reliability; stable errors that recur are invisible to consistency-based detectors, so future reference-free detectors need a recoverability signal."],"forward_implications":["FActScore-style verifiers can certify claims that directly contradict the passage they decomposed, so the decomposition step must be monitored as a source of error rather than trusted as bookkeeping.","Self-consistency sampling, a common reference-free hallucination detector, will miss DI-CC by design, because the injected content is stably recoverable and recurs across resamples.","Context-aware decoding can suppress DI-CC but introduces a severe faithfulness-completeness trade-off, including fabricated identities in coreference-heavy text, so it is not deployment-ready as implemented.","The mechanism's natural scope is narrow: it is rare under neutral instructions (0.2-0.4% of claims) and does not appear on FActScore's naturally occurring hallucinations, which generally fall in knowledge gaps rather than overridden correct beliefs.","There is a minimum scale threshold for the signal within the tested model family: 3B shows no transfer under either layer-selection criterion, while 7B does, suggesting the representation needed for the conflict mechanism only forms above some scale."],"supporting_citations":[{"why":"Supplies the NQ-Swap construction of classical context-memory conflict, the training data for the frozen probe.","marker":"Longpre et al., 2021"},{"why":"Defines FActScore's decompose-then-verify recipe and the real-world hallucination data used for the boundary replication attempt.","marker":"Min et al., 2023"},{"why":"Supplies the SelfCheckGPT self-consistency baseline that the paper shows fails to detect DI-CC.","marker":"Manakul et al., 2023"},{"why":"Supplies context-aware decoding, the classical mitigation the paper transfers to decomposition and finds only partially effective.","marker":"Shi et al., 2023"},{"why":"Supplies the variance-ratio layer-selection criterion used for the primary H0 result and the scale reanalysis.","marker":"Bürger et al., 2024"},{"why":"Supplies the layer-sweep probing methodology for locating conflict-related representations.","marker":"Pham et al., 2026"},{"why":"Supplies the context-memory conflict representation work underlying the probe-direction hypothesis.","marker":"Zhao et al., 2024"}],"fun_headline_variants":["Fact-checking pipeline can contradict its own source text","Decomposition can inject false claims that contradict the source","When a fact-checker’s split step betrays the original text","Parametric beliefs can override the text in fact-check pipelines","A decomposer’s memory can directly contradict the source passage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a linear probe trained on classical conflict data transferring to decomposition activations proves the two settings share a mechanism, rather than the probe tracking some task-general property such as unusual token positions or labeling artifacts; the below-chance result on Mistral-7B shows this transfer is not universal.","fun_headline_variants_meta":{"raw":{"variants":["Fact-checking pipeline can contradict its own source text","Decomposition can inject false claims that contradict the source","When a fact-checker’s split step betrays the original text","Parametric beliefs can override the text in fact-check pipelines","A decomposer’s memory can directly contradict the source passage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":3249,"prompt_tokens":1157,"completion_tokens":2092,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":773,"completion_tokens_details":{"reasoning_tokens":2009}},"tokens_in":773,"tokens_out":2092,"duration_ms":13849,"temperature":1.0,"reasoning_tokens":2009,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:37:44.119635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A control probe trained on NQ-Swap with conflict labels randomly permuted should not separate DI-CC positions from faithful decompositions; if a shuffled-label probe reaches AUC near 0.86 on the same Qwen2.5 activations, the claimed mechanism-specific transfer is not established. Alternatively, the paper's own Mistral-7B result, below-chance AUC under the same protocol, provides a case where the predicted transfer fails, so any complete account must explain why a genuine mechanism produces above-chance transfer in one model family and reliably wrong-direction transfer in another.","supporting_citations":[],"review_version":1}