{"id":"ccf3d291-75cf-4cad-b39e-8b0b2804a064","arxiv_id":"2506.10934","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A vector-based belief-update model, grounded in dynamic epistemic logic, predicts final block-weight beliefs in a collaborative task from dialogue friction.","lead":"This paper introduces a formal measure of epistemic friction, the resistance people show when new information clashes with what they already believe, and tests it on transcripts of a collaborative task. If the measure works, dialogue systems could better track when humans are hesitant, confused, or resistant to new evidence.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The update rule in Eq. (17) is inconsistent with its own worked examples: aligned red/green components would be shifted, not unchanged, so the reported RMSE rests on an unspecified or incorrect update.","rationale":"The reader identified the mismatch between Eq. (17) and the worked example as the weakest assumption, and this is indeed the most load-bearing concern. The central claim is that DEF, as formalized, predicts final belief states. The only quantitative evidence is the RMSE in Table 2, which is produced by applying Eq. (17) to hand-coded proposition vectors. If Eq. (17) cannot reproduce the paper's own illustrative outputs, then the experiments may have used an undocumented variant of the update, so the reported error does not validate the stated model. This is an internal inconsistency, not a question of consensus, and it is directly checkable by recomputation. I do not move the verdict to REJECT or UNVERDICTED because the issue may be a typographical or notational slip: the first worked example is consistent with Eq. (17), and the intended component-wise behavior might be recoverable from the accompanying text or code. The reader's CONDITIONAL verdict is therefore the right outcome: accept only when the update rule is corrected or clarified, code is released or the pipeline is re-run with the actual rule, and the revised RMSE is reported. I also note the paper's own admission that it does not compare against existing approaches and that the vectorization requires an isotropy property known to be inconsistent in LLMs; those are additional reasons the empirical support is provisional, but the formula inconsistency is the sharpest single point of failure.","tokens_in":12731,"tokens_out":5877,"duration_ms":57322,"concrete_test":"Recompute the second worked example exactly as written: φ_a=[10,10,20,0,0], φ_b=[10,-10,20,0,0], α=1, β=1, using Eq. (17) with standard cosine similarity. The output is [16.667, 3.333, 33.333, 0, 0], not [10, 3.333, 20, 0, 0]. If the authors' code or a corrected derivation reproduces the printed example, then Eq. (17) must be amended to the actual update rule and the experiments in Sec. 5.2 re-run under that rule; otherwise the reported RMSE cannot be attributed to the DEF update as described.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The empirical claim in Sec. 6 rests entirely on the vector-update operationalization in Eq. (17). That equation contradicts its own worked examples. For φ_a=[10,10,20,0,0] and φ_b=[10,-10,20,0,0] with α=β=1, Eq. (17) gives s=+1 and CosSim=2/3, so the updated vector is [16.667, 3.333, 33.333, 0, 0]; the paper reports [10, 3.333, 20, 0, 0]. The third example similarly yields [10, 4.523, 30.954, 0, 0], but the paper prints [10, 4.523, 20, 0, 0]. Under a single scalar coefficient multiplied component-wise by φ_b, aligned red and green components cannot remain unchanged. Thus either Eq. (17) was not the rule actually run, or the worked examples are wrong; the paper does not specify which, and no code is released. Since the conclusion that DEF 'effectively predicts' belief updates is supported only by the fitted RMSE in Table 2, and that RMSE is generated by an unspecified variant of Eq. (17), the central claim is not currently reproducible or testable. This is a correctness risk, not merely a missing baseline, because the mechanism being credited for prediction is not well defined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines dynamic epistemic friction (DEF) as resistance to belief revision in dialogue, formalizes it in Dynamic Epistemic Logic (DEL) terms and as vector-symbolic alignment, and evaluates a vector update rule on the Weights Task Dataset. Using hand-built R5 vectors for propositions and a ridge regressor from final belief vectors to ground-truth weights, it reports 2–3 g RMSE for most of four groups and concludes that DEF effectively predicts participant belief updates.","tokens_in":13094,"tokens_out":7270,"duration_ms":73414,"significance":"The conceptual move—quantifying belief-revision resistance via vector alignment and relating it to DEL—is interesting and could be useful for friction-aware dialogue systems. The paper is transparent about several limitations (isotropy assumptions, lack of direct competitors) and uses a public dataset, which are strengths. However, the empirical support is currently weak: the update rule is contradicted by its own worked examples, and the reported error is obtained after supervised regression and grid-searched coefficients rather than by a parameter-free DEF prediction. If the rule were corrected and proper baselines added, the framework could be a worthwhile contribution; in its present form the central claim is not reproducible from the manuscript.","major_comments":[{"comment":"Equation (17) is inconsistent with the worked examples in the same section. For φ_a=[10,10,20,0,0], φ_b=[10,-10,20,0,0], α=β=1, the equation gives φ'_a=[16.667,3.333,33.333,0,0], because CosSim=2/3 and s=+1, so both red and green components are reinforced. The paper instead prints [10,3.333,20,0,0]. The third example similarly gives [10,4.523,30.954,0,0] rather than [10,4.523,20,0,0]. If the experiments in §5.1 use Eq. (17), the implementation is different from the published rule; if they use the rule behind the printed examples, that rule is not specified. Either way, Table 2 cannot be reproduced or interpreted from the paper alone, and no code is released to resolve the ambiguity.","section":"§5, Eq. (17)"},{"comment":"The evaluation is supervised in a way that partly manufactures the result. The final extracted belief vector is run through a ridge regressor fit to the ground-truth final FBank, and α/β are chosen by grid search (with values 0.01..100 and k up to 4). The reported 2–3 g RMSE therefore measures the joint DEF+ridge+tuned-coefficient pipeline, not the predictive power of DEF itself. A ridge with five features trained on three dialogues can absorb systematic vectorization error; without a no-regression baseline (e.g., raw final vector vs. ground truth, or a constant predictor), the RMSE does not support the claim in §6 that DEF itself effectively predicts belief updates.","section":"§5.1, step 4"},{"comment":"The empirical basis is four dialogues, and no frictionless or naive baseline is implemented. The text claims a frictionless setting 'would involve a greater error than all models involving friction', but no such baseline appears in Table 2 or Fig. 2; low α/β values are not a frictionless condition because the update still uses Eq. (17) with nonzero coefficients. The conclusion that DEF outperforms frictionless updating is therefore not demonstrated. Additionally, group 4 has RMSE 7.7–17.5 g, so the 'most groups' claim in §5.2 should be qualified.","section":"§5.2, Table 2"},{"comment":"The claimed connection to DEL is not operationalized. Equation (17) is a heuristic vector update; the paper does not show that it corresponds to a DEL product update or to the evidence-bank transitions in (7), and axioms (5) are used only for direct component assignment. The R5 vectorization and the update rule are the actual mechanism being credited for prediction, but their relation to the formal DEL model is asserted rather than derived. If the contribution is the DEL-based theory, this gap needs to be filled; if it is the vector heuristic, the framing should be scaled back accordingly.","section":"§3–§5"}],"minor_comments":[{"comment":"There are numerous typos: 'isotopic' should be 'isotropic', 'alledges' should be 'allows', 'peformance' should be 'performance' (also in Fig. 2 captions), and §5 says '4 of of the 10 groups'.","section":"§6, Fig. 2, §5"},{"comment":"Equation (17) should define the Hadamard product ⊙ and clarify the scope of min; as written, min(β, α×s) could be read as min(β, α)×s.","section":"Eq. (17)"},{"comment":"Table 2 reports intervals like ±0.421 but does not state whether these are standard deviations, standard errors, or confidence intervals, nor over what (iterations, groups, or ridge fits) they are computed.","section":"Table 2"},{"comment":"The procedure says the focus participant is the one who speaks least, but also that 'which specific participant this is may vary across groups and is not further explicated here'; this should be made concrete for reproducibility.","section":"§5.1"},{"comment":"The representation of inequalities (e.g., yellow<40 as [0,0,0,0,40−U(0,1)]) should be justified; using a random draw in the vectorization introduces stochasticity that is not otherwise acknowledged in the update.","section":"§5"},{"comment":"The references are incomplete in places: Markowska et al. lacks year and venue, Ginzburg et al. contains a placeholder '1 others', and several arXiv preprints are cited without version or accession details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is readable and the theoretical framing is engaging. My main concern is that the empirical section, as written, would not allow a reader to reconstruct the update rule; I would want to see a corrected rule and released code before considering acceptance. The paper's fit to cs.CL is reasonable, but the strength of the conclusion should be brought in line with the small-scale, partly supervised evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe reader's take is fair, and the stress-test note lands. The paper's load-bearing claim—that DEF predicts belief updates—rests on Eq. (17), and that equation contradicts the paper's own worked examples. For the second example, the formula as written gives [16.667, 3.333, 33.333, 0, 0], not the printed [10, 3.333, 20, 0, 0]. The third is similarly off by the green component. This isn't a one-digit typo; it means the implemented update, whatever it was, is not the one written. No code is released, so the reported RMSEs in Table 2 come from an unspecified variant. That is a correctness problem, not just a missing baseline.\n\nWhat the paper does well: the notion of epistemic friction as resistance to belief revision is a genuinely useful lens for dialogue, and the vector operationalization with cosine similarity and HRR-style vectors is a clean way to make it computable. The connection to evidence-based DEL is appropriate, and the Weights Task dataset is a decent testbed. The paper is clearly written and, to its credit, openly notes the isotropy limitation and admits it did not compare against other methods.\n\nThe soft spots beyond the equation are serious. The ridge regressor maps the final vector to ground truth, so the 2–3g RMSE partly measures that fit, not the model's predictive power. The friction coefficients α and β are tuned by grid search on the held-out groups, adding another fitted element. The evaluation uses only 4 dialogues. And the claim that 'without friction, error is greater' is not backed by an implemented no-friction baseline; the paper asserts it.\n\nWho is this for? Researchers working on common-ground tracking, human-AI collaboration, or vector-symbolic models of dialogue. There is a real idea here. But the empirical demonstration as it stands is not reproducible or verifiable. A serious referee should send it back for major revision: fix or derive the update rule, release code, add a true no-friction baseline, and clarify what the ridge is allowed to learn. I would accept it for peer review for that reason, but I would not cite it in this form.","headline":"A promising conceptual framework, undermined by an update rule that contradicts its own worked examples and an evaluation that fits rather than predicts.","tokens_in":13583,"tokens_out":3499,"would_cite":false,"duration_ms":35564,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that dynamic epistemic friction, quantified as vector misalignment, predicts how participants revise beliefs in a collaborative dialogue to within a few grams.","keywords":["dynamic epistemic friction","belief revision","dynamic epistemic logic","vector symbolic architectures","holographic reduced representations","common ground","dialogue state tracking","Weights Task Dataset"],"falsifier":"Execute equation (17) as written on the paper's own second worked example: with $\\vec{\\phi}_a=[10,10,20,0,0]$, $\\vec{\\phi}_b=[10,-10,20,0,0]$, and $\\alpha=\\beta=1$, the formula scales every nonzero component of $\\vec{\\phi}_b$ by the same factor and adds it to $\\vec{\\phi}_a$, so the red and green components should move away from their starting values. The paper reports the updated vector as $[10,3.333,20,0,0]$, leaving red and green exactly unchanged. Recomputing this one line settles whether the published update rule is the rule that produced the reported predictions.","tokens_in":12554,"feed_emoji":"⚖️","tokens_out":16836,"duration_ms":158125,"temperature":0.7,"pith_summary":"This paper tries to establish that dynamic epistemic friction—the resistance a person feels when new information conflicts with, or sits far from, what they already believe—is a measurable quantity that can be used to predict how beliefs change in dialogue. The authors formalize friction inside Dynamic Epistemic Logic as persistent misalignment between a belief state and an incoming proposition, then encode belief states and assertions as five-dimensional vectors so that alignment becomes a geometric similarity. On four annotated sessions of a collaborative block-weighing task, they feed a participant's belief vector through an update rule with two tuned friction coefficients as each interlocutor utterance arrives; after a fitted linear map, the final vector predicts the true block weights to within 2-3 grams RMSE for three of the four groups. The intended consequence is that resistance to belief revision is not an obstacle to be smoothed away but a signal that dialogue systems can measure, tune, and exploit.","feed_headline":"Friction metric predicts final beliefs to within 2-3 grams","feed_subtitle":"A cosine-similarity update rule with tuned coefficients tracks how participants revise beliefs in a collaborative task.","key_machinery":"The engine is a vector-space operationalization of Dynamic Epistemic Logic. Belief states and propositions live in $\\mathbb{R}^5$ with components ordered [red, blue, green, purple, yellow]; an assertion such as red=10 and blue=10 is written $[10,10,0,0,0]$, and a denial of a claimed weight contributes a negative entry. The load-bearing update rule is $$\\vec{\\phi}'_a = \\vec{\\phi}_a + \\min\\{\\$\\beta$,\\$\\alpha$\\,s\\}\\cdot \\mathrm{CosSim}(\\vec{\\phi}_a,\\vec{\\phi}_b)\\odot \\vec{\\phi}_b,\\qquad s=\\mathrm{sgn}(\\vec{\\phi}_a\\cdot \\vec{\\phi}_b),$$ where $\\odot$ is elementwise multiplication, $\\alpha$ is the strength of the friction force, and $\\beta$ caps how much an aligned assertion can reinforce an existing belief. Friction itself is defined as one minus the cosine-similarity alignment between the belief vector and the proposition-plus-evidence vector, and the QBank-to-EBank-to-FBank structure from evidence-based DEL supplies the vocabulary of where propositions sit as evidence accumulates. The coefficients $\\alpha$ and $\\beta$ are the degrees of freedom that make the model resist or accelerate belief change, and the paper grid-searches them, finding $\\alpha=5$, $\\beta=2$ best on the Weights Task data.","core_discovery":"The paper's central claim is that epistemic friction can be turned from a metaphor into a computational quantity: define friction as one minus the alignment between an agent's belief vector and the incoming proposition plus its evidence, then update the belief vector by a step whose size and sign depend on that alignment. Concretely, the model predicts the final belief state of the least-talkative participant in each group of the Weights Task by starting from a uniform initial vector, applying the update rule to every interlocutor assertion, and regressing the resulting vector onto the known true weights $[10,10,20,30,50]$. The reported root-mean-squared error of 2-3 grams for most groups is offered as evidence that the resistance profile the model computes tracks the way human participants actually revise their beliefs, and that the two friction coefficients control a trade-off between naive acceptance and rigid refusal of new information.","pith_inferences":["Editorial extension: because the final belief vector is passed through a fitted ridge regressor before scoring, the experiment measures whether DEF dynamics correlate with the final belief state, not whether the update rule alone produces gram-accurate weights; a control that feeds only utterance counts per block into the same regressor would show how much of the fit is friction versus mere dialog","Editorial extension: the model assumes utterance order matters, so shuffling the utterances in a group's dialogue should change the predicted final state; if predictions are order-insensitive, the 'dynamic' part of DEF is doing little work beyond counting evidence per block.","Editorial extension: substituting learned LLM embeddings for the hand-built [red, blue, green, purple, yellow] vectors would test the paper's admitted isotropy caveat; if the update rule fails with real embeddings, the friction concept may still hold but the hand-built representation, not the concept, is what carried the empirical result."],"forward_implications":["Belief revision in dialogue becomes tunable: a system can decide how much resistance to apply to each incoming assertion rather than treating every accepted utterance as a flat update.","The optimal coefficients show a 'just enough friction' regime: too little resistance lets beliefs chase incorrect intermediate positions, too much makes them rigid, so the two coefficients act as regularizers on prediction error.","The same update machinery extends beyond collaborative tasks to adversarial ones such as Diplomacy, where a spike in friction between aligned agents could flag the moment one agent starts deceiving another.","The distance between a participant's current belief vector and a proposed update can feed a classifier that predicts the participant's next move—acceptance, counterargument, or a clarifying question—using the gap the paper equates with friction."],"supporting_citations":[{"why":"Supplies the logical framework of evidence-based DEL that the paper extends, defining belief updates via product updates and evidence strength.","marker":"Van Benthem and Pacuit (2011)"},{"why":"Provides the product-update formalism and the Saying/Seeing/Acting-is-Believing axioms used to infer participants' beliefs from utterances.","marker":"(Bolander, 2014)"},{"why":"Introduces the QBank/EBank/FBank common-ground structure and the block-weight vector ordering the paper's vectorization relies on.","marker":"(Khebour et al., 2024b)"},{"why":"Provides the Weights Task Dataset—the dialogues, annotations, and ground-truth block weights used in the prediction experiments.","marker":"(Khebour et al., 2024a)"},{"why":"Supplies the TRACE annotations of asserted beliefs from speech, gesture, gaze, and action that define the belief states being updated.","marker":"(VanderHoeven et al., 2025)"},{"why":"Establishes Holographic Reduced Representations, the vector-symbolic basis for encoding propositions so that alignment can be measured by cosine similarity.","marker":"(Plate, 1995)"}],"fun_headline_variants":["Friction model pins beliefs to within 2-3 grams","Belief revision resistance quantified in LLM dialogues","Predicting belief updates via epistemic friction","Dialogue friction: a metric for belief resistance","From metaphor to metric: friction predicts belief shifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole empirical result rests on the assumption that writing beliefs as five hand-assigned numbers and updating them with the cosine-scaled rule (17) is a faithful implementation of the logical belief revision that DEL describes; if the vectorization and the rule drift from the logic, the error bars measure the linear regressor, not epistemic friction.","fun_headline_variants_meta":{"raw":{"variants":["Friction model pins beliefs to within 2-3 grams","Belief revision resistance quantified in LLM dialogues","Predicting belief updates via epistemic friction","Dialogue friction: a metric for belief resistance","From metaphor to metric: friction predicts belief shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1422,"prompt_tokens":881,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":497,"tokens_out":541,"duration_ms":7165,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:13:22.099518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Execute equation (17) as written on the paper's own second worked example: with $\\vec{\\phi}_a=[10,10,20,0,0]$, $\\vec{\\phi}_b=[10,-10,20,0,0]$, and $\\alpha=\\beta=1$, the formula scales every nonzero component of $\\vec{\\phi}_b$ by the same factor and adds it to $\\vec{\\phi}_a$, so the red and green components should move away from their starting values. The paper reports the updated vector as $[10,3.333,20,0,0]$, leaving red and green exactly unchanged. Recomputing this one line settles whether the published update rule is the rule that produced the reported predictions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the logical framework of evidence-based DEL that the paper extends, defining belief updates via product updates and evidence strength."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the product-update formalism and the Saying/Seeing/Acting-is-Believing axioms used to infer participants' beliefs from utterances."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TRACE annotations of asserted beliefs from speech, gesture, gaze, and action that define the belief states being updated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes Holographic Reduced Representations, the vector-symbolic basis for encoding propositions so that alignment can be measured by cosine similarity."}],"review_version":1}