{"id":"69ec77e3-f7f7-46b5-a5a6-ff7a8ac97b4f","arxiv_id":"2601.11429","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Across six hand-picked relations and four instruction-tuned models, a relation's measured linearity (Δcos) positively correlates with hallucination rates on synthetic unknown entities (r≈.78–.82), though p-values mostly exceed .05.","lead":"Language models hallucinate invented attributes when asked about made-up entities; this paper links that tendency to how linearly the model stores a relation. Across four models and six relations, hallucination rate tracks a representation-space linearity score, but the sample is tiny and the statistics are borderline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Δcos proxy is not separated from answer-space priors; Appendix G's control is circular, so the r≈.8 correlation may reflect output concentration rather than the proposed storage-format mechanism.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Δcos is only a practical proxy, and the paper's own limitations concede that answer-space priors may confound the relationship. My stress-test sharpens this by pointing out that Appendix G's concentration controls are computed from hallucinated outputs, making the control circular rather than an independent test. The statistical evidence is otherwise transparent—Wilson intervals, exact permutation tests, LOO ranges, and a regex baseline are all good practices—but none of these address the fundamental proxy-validity issue. The abstract/text discrepancy over 15 vs six relations and r-ranges further suggests the reported correlation is sample-dependent. Because the central claim is causal about storage format, and the data are correlational with an unresolved confound, the appropriate verdict remains CONDITIONAL exactly as the reader concluded. No change to the verdict is needed, but the paper should add the proposed independent prior-concentration control and layer sensitivity analysis before the claim can be accepted as mechanistic.","tokens_in":13442,"tokens_out":4927,"duration_ms":52436,"concrete_test":"Recompute the main correlation using the 15 relations from the abstract (or add at least 9 new relations matched for entity type and answer-set size). For each relation, estimate an output-prior concentration covariate that is independent of the model's generations: e.g., the normalized entropy of gold-object strings in the pretraining corpus, or the top-10 token probability mass of a frozen base model given the question template with a blank/unknown subject. Then compute the within-model partial correlation between Δcos (Eq. 3) and hallucination rate controlling for this covariate on n=15 relations. If the partial r is no longer significantly positive, or if a layer sweep changes country_language's Δcos ranking, the headline correlation does not support the storage-abstractness mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that Δcos measures how abstractly a relation is stored, independently of how concentrated the relation's plausible answer space is. The paper's own Limitations admits that answer-space entropy can affect both hallucination propensity and representation clustering that influences Δcos, and only Appendix G attempts to control it. But the Appendix G proxies (Top-1 share and normalized entropy) are computed from the model's own hallucinated outputs—the very behavior being predicted—so the control is partly circular and cannot establish an independent prior. The six relations also conflate relation semantics with answer-set size: instrument/sport have small closed answer sets, father/CEO are long-tailed open sets, and country_language rests on only npairs=24/ntest=6 LRE triples (Table 4). A relation with concentrated plausible answers will yield high hallucination rates (any plausible guess is easy) and may also yield high Δcos because many object representations lie near a single direction; both effects could be driven by that concentration, not by storage abstractness. Given that the abstract reports 15 relations with r∈[.58,.84] while the paper text uses six relations with r∈[.78,.82], the strength of the claimed predictor is sensitive to the relation sample. The data as presented therefore do not yet separate the proposed storage-format mechanism from an output-space-prior confound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes that the degree of relational linearity in LLM representations—measured by Δcos on natural triples—predicts whether the model hallucinates an object for a synthetic unknown subject versus refuses to answer. The authors introduce SyntHal, a dataset of 6000 synthetic entities across six relations, and measure hallucination rates for four instruction-tuned LLMs. They report a strong positive correlation (r≈.78–.82) between relation-level Δcos and hallucination rate, and interpret this as evidence that linearly stored relations make knowledge self-assessment harder. The paper also includes a natural-triple analysis (LRE) showing the opposite correlation, explained as a relation difficulty effect.","tokens_in":13704,"tokens_out":5933,"duration_ms":57242,"significance":"If the central claim holds, the paper identifies a representational property—relational linearity—as a predictor of hallucination behavior, with potential implications for knowledge self-assessment and targeted mitigation. The measurement pipeline is careful and reproducible in several respects: Wilson confidence intervals for hallucination rates, exact permutation p-values, leave-one-relation-out ranges, a regex judge baseline with 96.4% agreement, and human validation of 200 judge labels. The use of external natural triples for linearity and separate synthetic entities for behavior reduces direct circularity. However, the evidence is correlational and rests on only six hand-picked relations, and the proposed storage-format mechanism is not cleanly separated from output-space priors. The paper is transparent about these limitations, but they are load-bearing for the interpretation.","major_comments":[{"comment":"The central evidence is a correlation over n=6 hand-picked relations. Two-sided exact permutation p-values are all >0.05 (0.071, 0.064, 0.089, 0.086); only one-sided p-values are below .05. The Fisher combined p=0.0085 combines non-independent samples (the same six relations across four models) and is not a valid significance test. Because relations were selected to maximize linearity spread, the correlation may be inflated. Please add more relations (e.g., all LRE relations with sufficient triples), report confidence intervals for r, and use a test that accounts for relation identity.","section":"§4, Figure 1, Table 6"},{"comment":"The output-space concentration control is partly circular: Top-1 share and normalized entropy are computed from the model's own hallucinated outputs—the very behavior being predicted. If answer spaces are concentrated, both high hallucination rate and high Δcos can result without storage-format abstraction being the cause. Please provide an independent measure of answer-space concentration (e.g., from a corpus or from the model's prior over known subjects) or use interventions (patching/steering) that hold answer-space entropy fixed. As it stands, the data do not separate the proposed storage-format mechanism from an output-space-prior confound.","section":"Appendix G, Limitations"},{"comment":"The abstract states SyntHal contains 15 relations and reports r∈[.58,.84]; the main text (Abstract, §2.2, Figure 1) says six relations and r∈[.78,.82]. These are different claims. The manuscript must be reconciled—readers cannot assess which result is being claimed.","section":"Abstract vs. main text"},{"comment":"The linearity proxy uses a translation-only probe (W=I) and fixed layers ℓs=⌊L/2⌋, ℓo=L−2, with no robustness analysis. If Δcos is sensitive to these choices, the correlation could be a probe artifact. Please report Δcos across a range of layers and with an unconstrained affine probe (W free) for the same relations to show the main result is stable.","section":"§3, Eq. (1)–(2), Appendix D"}],"minor_comments":[{"comment":"The x-axis label reads 'LRE cosine improvement ( cos)' but should be 'Δcos' for consistency with Eq. (3).","section":"Figure 1"},{"comment":"The entry 'Lindsey and 1 others. 2025' is unprofessional; either list the full author list or use standard 'et al.' formatting.","section":"References"},{"comment":"The relation name 'father’s first name' is awkward as a relation label; consider a canonical form such as 'father_first_name'.","section":"Table 2"},{"comment":"The statement that ignoring covariance is 'slightly conservative' is correct only because the covariance is typically positive; this could be made explicit.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The paper is a careful empirical study with a transparent measurement pipeline, but the central claim rests on a small, hand-picked set of relations and a confound that is not fully controlled. The abstract inconsistency (15 vs 6 relations) is serious and must be fixed. I recommend major revision rather than rejection, as the core hypothesis is interesting and the authors already acknowledge the main limitations; a revised version with more relations and an independent answer-space control could make the case convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before believing its headline result. The core claim is clean: for synthetic unknown entities, the linearity of a relation (measured by Δcos on natural triples) predicts whether a model hallucinates an object or refuses, with r≈.78–.82 across four 7B models. If true, it gives a representation-level handle on hallucination. The paper builds on Hernandez et al.'s LRE work and does something genuinely new: it connects the storage format of a relation to the model's ability to know what it doesn't know.\n\nWhat's good: the measurement is unusually careful. The LLM judge was validated on 200 random examples with 100% agreement, and a deterministic regex baseline hits 96.4% agreement on 24,000 outputs. They report Wilson CIs, exact permutation p-values, leave-one-relation-out ranges, and weighted correlations. The limitations section is honest — they admit the translation-only probe is a proxy, the analysis is correlational, and answer-space entropy is a confound.\n\nNow the soft spots. The first is the n=6 problem. Three of four within-model p-values are above .05; the combined Fisher p relies on non-independent samples. The abstract says 15 relations with r∈[.58,.84], but the paper text uses six relations with r∈[.78,.82]. That's not a minor typo — it changes the strength of the claimed predictor, and the discrepancy needs an explanation. Second, the Δcos proxy may reflect output-space concentration rather than storage abstractness. A relation like sport has a small answer set, so any plausible guess is easy and the object representations may cluster along a single direction; that alone could drive both high hallucination rates and high Δcos. Appendix G tries to control for this, but the proxies are computed from the model's own hallucinated outputs — the very behavior being predicted. That control is partly circular. The sign reversal on natural triples gets a plausible post-hoc explanation, but it shows the relationship is not universal.\n\nIs the central argument load-bearing? It's a reasonable hypothesis, and the evidence is suggestive but not conclusive. The paper deserves a serious referee, because the question matters and the methodology is mostly transparent. I'd send it to review, but with a clear request: expand the relation set (or at least fix the abstract), add sensitivity analyses over layers and probe constraints, and redo the output-prior control using external answer distributions, not the model's own outputs. For a reading group, I'd bring it as a cautionary example of how a clean idea can outrun its statistical support.\n\nRecommendation: engage with it — but treat the r≈.8 as an upper bound, not a settled fact.","headline":"A plausible mechanism linking relation linearity to hallucination, with a careful evaluation pipeline, but the evidence rests on only six relations, an abstract/text mismatch, and a circular control for output priors.","tokens_in":14279,"tokens_out":1275,"would_cite":true,"duration_ms":15321,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Relational linearity predicts whether a language model hallucinates or refuses when asked about an invented entity.","keywords":["hallucination","relational linearity","Δcos","language models","knowledge self-assessment","synthetic entities","refusal behavior"],"falsifier":"Intervention test: take a nonlinear relation on which models mostly refuse for unknown subjects (e.g., father's first name), add the relation's average difference vector to the subject representation at the probed layer, and decode greedily. If refusals persist, linearity is not the causal mechanism; if they turn into confident hallucinations, the translation-vector story is confirmed. A complementary check is to recompute Δcos with a full affine map (allowing W not equal to identity) and see whether the r≈0.8 correlation survives.","tokens_in":63,"feed_emoji":"🤖","tokens_out":8961,"duration_ms":153563,"temperature":0.7,"pith_summary":"The paper claims that a language model's tendency to fabricate an answer for an entity it has never seen is predictable from how linearly the model stores the underlying relation. Using six relations probed with 6,000 synthetic, unknown subjects, four instruction-tuned models hallucinated—gave a concrete value instead of refusing—far more often for relations whose subject-to-object mapping is well captured by a single added vector, with correlations between the linearity score and hallucination rate of about r = 0.78–0.82. The proposed reason is that a linear relation supplies a reusable translation that produces a plausible object even when no specific triple was learned, so the model cannot tell that it does not know; nonlinear relations require more instance-specific storage, which makes refusal easier. If this is right, part of the blame for hallucinations shifts from missing knowledge to the format in which knowledge is stored, and mitigation can target high-linearity relations with explicit known-fact checks.","feed_headline":"Linearity predicts when LLMs hallucinate","feed_subtitle":"Relations stored as one vector shift make models answer invented entities instead of refusing—r≈0.8 across four models.","key_machinery":"The load-bearing measure is Δcos, computed with a deliberately simple translation-only probe: for each model and relation, subject and object hidden states are extracted from natural triples at fixed layers (mid-layer for subjects, third-from-last for objects); the relation direction d̄r is the mean object-minus-subject vector on a training split; the object is predicted as s + d̄r on held-out subjects; and linearity is the expected cosine improvement over using the subject itself. A high Δcos means one shared direction reproduces many triples—the relation is stored 'abstractly.' That same direction is what lets the model generate a plausible object for a never-seen subject, which is the pro","core_discovery":"On the paper's own terms, the discovery is that relational linearity is a strong relation-level predictor of hallucination on prompts where the subject is unknown by construction. The authors generate synthetic entities across six relations, ask four instruction-tuned models for the relation's object, and label each answer as refusal or hallucination. Separately, they measure linearity with Δcos: on natural subject–object triples, they take the average subject-to-object difference vector, add it to held-out subject representations, and record the improvement in cosine similarity to the true object. Hallucination rate rises with Δcos for all four models (Pearson r between .78 and .82), and th","pith_inferences":["A natural extension is an intervention: add the learned relation direction to an unknown subject's representation and check whether refusals turn into hallucinations; this would convert the correlation into a causal test.","The authors' output-concentration controls are suggestive but not definitive; a matched-pair design that pairs relations with similar answer entropy but different Δcos would cleanly separate the storage-format story from a prior-concentration story.","The reversal on natural triples hints that linearity partly tracks training frequency: common relations become linear and accurate. Controlling for pretraining counts across relations would tell whether linearity adds predictive power beyond frequency.","If the mechanism generalizes, a practical diagnostic could rank relations by Δcos on a small probe set and use it to decide where refusal guardrails are needed before deployment."],"forward_implications":["Hallucination risk for a relation can be estimated from representation geometry on ordinary known triples, before the model is ever asked about an unknown entity.","For highly linear relations, a confident fabrication does not require the model to 'know' anything: the translation vector supplies an object automatically, so refusal behavior is not a straightforward honesty signal.","Mitigations should add explicit known-triple verification for linear relations; supplementing abstract representations with evidence markers is the concrete target the authors propose.","Linearity is a continuum, not a binary: intermediately linear relations show intermediate hallucination rates, so interventions should be graded.","On natural (known) triples the correlation reverses—more linear relations are more accurate—so the effect is specific to genuinely unknown subjects and evaluations must keep the two regimes separate."],"fun_headline_variants":["Linear relation storage predicts LLM hallucinations","Hallucination risk scales with relational linearity","LLMs invent answers when relations are linear","Relational linearity foretells LLM hallucination vs refusal"],"cache_read_input_tokens":15488,"weakest_assumption_plain":"The load-bearing premise is that Δcos—how much one averaged subject-to-object difference vector improves object prediction on held-out natural triples—captures how abstractly the model stores the relation; if Δcos mostly tracks answer-output concentration, layer choice, or object frequency instead, the correlation would not support the proposed storage-format mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Linear relation storage predicts LLM hallucinations","Hallucination risk scales with relational linearity","LLMs invent answers when relations are linear","Relational linearity foretells LLM hallucination vs refusal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000688,"raw_usage":{"total_tokens":2947,"prompt_tokens":726,"completion_tokens":2221,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2161}},"tokens_in":470,"tokens_out":2221,"duration_ms":20551,"temperature":1.0,"reasoning_tokens":2161,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:59:48.762095+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Intervention test: take a nonlinear relation on which models mostly refuse for unknown subjects (e.g., father's first name), add the relation's average difference vector to the subject representation at the probed layer, and decode greedily. If refusals persist, linearity is not the causal mechanism; if they turn into confident hallucinations, the translation-vector story is confirmed. A complementary check is to recompute Δcos with a full affine map (allowing W not equal to identity) and see whether the r≈0.8 correlation survives.","supporting_citations":[],"review_version":1}