{"id":"ec08976a-b180-4626-8eae-5111fd43ed0f","arxiv_id":"2506.05196","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"LPMT combines ensemble bidirectional diffusion, reciprocal-neighborhood distribution embeddings, and a Sinkhorn-Wasserstein distance, claiming state-of-the-art re-ranking on ROxf and RPar.","lead":"LPMT is a re-ranking method that turns each image into a probability distribution and measures retrieval distance as the cheapest thermodynamic transition cost between distributions, reporting mAP gains on Oxford5k and Paris6k. The paper is a candidate read for anyone who wants to know whether optimal-transport-style distances can replace classic diffusion propagation in the final ranking stage of image search systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multi-state TMT path is the load-bearing claim, but Eq. (20) only defines a locally constrained shortest path; the paper gives no path-optimization algorithm and Table 6 never compares against direct W1, so the reported gains do not establish the mechanism.","rationale":"The reader correctly identifies the TMT mechanism as the most fragile part of the paper, but their specific mathematical reason — that metricity of W1 collapses the path minimization — is invalid because Eq. (20) restricts π to locally reachable states, so the direct W1 edge may not be in the feasible set. The valid core of the concern is that the paper never demonstrates that the multi-state path matters: no algorithm for path optimization is provided, no code is released, and the ablations do not isolate the path contribution. The reported gains could come entirely from using W1 as a distance between the LSE distributions, with the thermodynamic multi-state narrative being irrelevant. Because the paper's central claim is precisely that the long-term multi-state process improves retrieval, this missing evidence is load-bearing. The proposed concrete test would settle the issue: if direct W1 matches the path-based d', the mechanism is vacuous and the paper should be rejected or reframed; if it does not, the authors need to supply the path algorithm and a corresponding ablation, which is a solvable revision. Hence CONDITIONAL rather than outright REJECT is appropriate, since the flaw is one of missing evidence and specification rather than a proven internal contradiction in Eq. (20). The other noted issues (BCD gradient typo, no code, hyperparameters tuned on test benchmarks) reinforce the need for a conditional verdict pending these checks.","tokens_in":30839,"tokens_out":10851,"duration_ms":128309,"concrete_test":"Implement two versions of d' on ROxf with R-GeM: (A) direct W1 between pi and pj with no intermediate states (K=0); (B) locally constrained shortest path over the k-NN graph G, where edge weight between local neighbors is W1 (e.g., Floyd-Warshall O(n^3) with k from Fig. 5a). Compare the retrieval mAP. If A and B produce identical rankings, the multi-state transition contributes nothing and the central claim collapses; if B is better, the mechanism is real, but the authors must then report the exact path-optimization algorithm to make the claim reproducible and to justify the claimed O(n^3) complexity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the long-term multi-state thermodynamic transition contributes to retrieval (Conclusion, Eq. (21)). The reader's collapse argument is not itself correct: Eq. (20) minimizes over paths π that are explicitly constrained to step through local regions ('each state is reachable from its predecessor only within the local region'), so the direct W1 edge is not necessarily feasible and the minimum over the restricted path set is generally larger than W1(pi,pj). However, a load-bearing concern survives in a different form: the paper never specifies how the path optimization over π is performed (no algorithm is given for the constrained shortest path; Appendix B.3 only discretizes the pairwise W1 via Sinkhorn), and the TMT ablation (Table 6) compares against Cosine/Euclidean/Jaccard/Total Variation but not against direct W1 (K=0) or an unconstrained path. If the implementation simply computes W1(pi,pj) for all pairs, the multi-state mechanism is vacuous and the reported mAP gains are attributable to the LSE + W1 pipeline rather than to the thermodynamic story. The proof in Appendix B.2 only establishes that one stage's cost equals W1 for a fixed path; it does not establish that path minimization changes the outcome. Thus the central claim is currently unsupported by the evidence provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LPMT, a re-ranking framework for instance retrieval comprising three components: Bidirectional Collaborative Diffusion (BCD) to fuse multiple affinity graphs, Locality State Embedding (LSE) to represent each image as a probability distribution over its local neighborhood, and Thermodynamic Markovian Transition (TMT) to compute a distance between these distributions via a multi-state transition flow. The final distance is a weighted combination of the original Euclidean distance and the TMT cost. Experiments on ROxf/RPar and on content-based retrieval datasets report consistent mAP improvements over a wide range of baselines, including substantial gains over the previous state of the art (e.g., R-GeM ROxf medium 84.7 vs. CAS 80.7).","tokens_in":31092,"tokens_out":7731,"duration_ms":82393,"significance":"If the multi-state thermodynamic transition were the source of the reported gains, this would be a meaningful contribution to manifold-based re-ranking: the empirical improvements are large, consistent across six feature extractors, and the paper includes a broad set of baselines. The paper also provides a formal connection between the flow cost and the W1 distance. However, the central claim about the multi-state path is not operationalized: the paper never defines the feasible path set, gives no algorithm for the constrained path optimization, and the TMT ablation omits the direct W1 baseline. As presented, the evidence does not isolate the contribution of the thermodynamic path mechanism, which is the paper's main novelty.","major_comments":[{"comment":"The paper never specifies the set of feasible transition paths π: it states that each state must be reachable from its predecessor 'within the local region', but it does not define what 'local region' means for the distributions {p_i} or how the constrained shortest path in Eq. (20) is computed. Appendix B.3 only provides a Sinkhorn iteration for pairwise W1. Table 6 compares TMT against Cosine, Euclidean, Jaccard, and Total Variation distances, but omits the direct W1 baseline (K=0 or a single-edge path) and an unconstrained-path baseline. As a result, the reported gains do not establish that the multi-state transition contributes; they could be entirely due to the LSE plus W1 pipeline. This is the load-bearing claim of the paper and must be supported by an explicit path-optimization algorithm and the missing baselines.","section":"§4.3, Eq. (20)"},{"comment":"The proof shows only that for a fixed path π, the minimum flow cost over transition rates equals the sum of W1 distances along the path; it does not show that the minimum over paths is different from the direct distance W1(p_i, p_j), nor does it enforce any local restriction on the transport plan. In the constructive part (Eqs. B.21–B.28), mass is moved directly between arbitrary vertices r_m and s_m over edges with cost d(r_m, s_m), which is exactly the unconstrained optimal transport problem. If the intended 'local region' constraint applies to these transport edges, the proof does not model it; if it applies only to the sequence of intermediate states, that constraint and its effect on the optimum need to be stated formally.","section":"Appendix B.2, proof of Eq. (B.8)"},{"comment":"The gradient of J in Eq. (7) is inconsistent with the quadratic form in Eq. (6) and with the correct derivative in Appendix A.1 (Eq. A.9). For symmetric ar{S}^v, the derivative of vec(F)^\top (I - ar{S}^v) vec(F) is 2(I - ar{S}^v)vec(F), not (2I - ar{S}^v)vec(F). The closed-form solution in Eq. (8) corresponds to the correct gradient, so this appears to be a typographical error, but it should be corrected because the text presents Eq. (7) as the basis for Eq. (8).","section":"§4.1, Eq. (7)"},{"comment":"The statement that 'an additional power term can be applied to the distance in Eq. (18)' is never formalized; no definition, default value, or experimental setting is given. Either specify the term (e.g., d(r,s)^p) and its value in the experiments, or remove the sentence, because the reader cannot determine what was actually evaluated.","section":"§4.3, text after Eq. (18)"}],"minor_comments":[{"comment":"'Temperate states' appears to be a typo for 'temporary states' or 'intermediate states'.","section":"§4.3, Eq. (18) and surrounding text"},{"comment":"The header 'k-reciprcal' should be 'k-reciprocal'.","section":"Table 5"},{"comment":"The initialization of P(0) and R(0) is stated in a single line, but R(0) is not subsequently used in the loop; clarify the notation and the stopping criterion.","section":"Appendix A.3, Algorithm 2"},{"comment":"The paper does not report the TMT-specific hyperparameters (e.g., τ, K, the Sinkhorn regularization ε, and the number of path states), even though the sensitivity analysis in Fig. 5 covers only k1, k2, κ, θ, μ, and σ. Please add these details to make the experiments reproducible.","section":"§5.1, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper's advertised novelty is the multi-state thermodynamic transition, but the current manuscript does not demonstrate that this mechanism is active: the path-optimization procedure is unspecified and the direct-W1 baseline is missing. The authors should be asked to report the exact algorithm used for Eq. (20) and to compare against single-edge W1 and an unconstrained shortest path. In addition, the relationship to the authors' prior CAS work (Luo et al., 2024) should be clarified, since BCD appears to be a multi-graph extension of the CAS diffusion. The empirical results are strong and the paper is well structured, so the issues are fixable within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The experiments are extensive and the gains are consistent: 3–4 mAP over CAS on ROxf with R-GeM, and similar patterns across DOLG, CVNet, MAC, R-MAC, DELG, and SENet. If those numbers replicate, the BCD+LSE+W1 pipeline is useful for instance retrieval re-ranking. The second thing: the paper's central story about a multi-state thermodynamic transition is not supported by the evidence as written. The reader's collapse argument is close but not exact—Eq. (20) does constrain the path to local states, so the direct W1 edge is not necessarily feasible—but the stress-test note has the correct fix: the paper never gives an algorithm for the path optimization over π, and Table 6 compares against Cosine/Euclidean/Jaccard/Total Variation, not against direct W1 or an unconstrained one-step path. If the implementation simply computes W1 for every pair, the multi-state process contributes nothing and the gains come from the LSE representation plus W1. If it computes a constrained graph shortest path, that is a different algorithm that needs to be described and ablated. Either way, the load-bearing claim about long-range thermodynamic transitions is untested.\n\nCredit where due: the BCD derivation in the appendix is careful, and the closed-form solution and iterative scheme are consistent with each other (the main-text Eq. (7) has a factor-of-two typo versus the appendix—minor, not load-bearing). The W1 equivalence proof for a single stage is fine. LSE is a sensible k-reciprocal embedding. The paper covers many baselines and backbones, which is rare and valuable.\n\nSoft spots beyond the TMT gap: no code is provided, hyperparameters are tuned on the test benchmarks, and novelty is limited—BCD resembles automatic ensemble diffusion and LSE resembles k-reciprocal encoding. The empirical gains could still hold, but the current manuscript overclaims what TMT does.\n\nWho this is for: researchers in diffusion-based re-ranking. It deserves a serious referee because the empirical contribution is substantive even if the theoretical framing needs rewriting. My recommendation: send to peer review, but expect major revision—reframe TMT honestly (e.g., as a Sinkhorn-Wasserstein distance on locality embeddings), add the missing direct-W1 ablation and path-optimization details, and release code.","headline":"Strong empirical re-ranking results, but the thermodynamic multi-state story is untested as written; ship to review with major revision.","tokens_in":31727,"tokens_out":4463,"would_cite":false,"duration_ms":51216,"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":"The paper proposes LPMT, a manifold-aware distance measured as the minimum transition flow cost between per-image probability distributions, and reports consistent gains in instance retrieval re-ranking.","keywords":["instance retrieval","re-ranking","manifold ranking","diffusion","optimal transport","Wasserstein distance","thermodynamic transition","locality preserving"],"falsifier":"Compute $d'(i,j)$ from Eq. (20) with one intermediate state and with many for a sample of image pairs; since $W_1$ is a metric, the values coincide, which would show the multi-state transition contributes nothing and any gains come from LSE, BCD, or the Euclidean blend. Alternatively, check the released implementation to see whether $W_1$ is restricted to transitions along local graph edges; if it is not, the computed distance reduces to the direct earth-mover distance.","tokens_in":30560,"feed_emoji":"🖼️","tokens_out":9289,"duration_ms":95480,"temperature":0.7,"pith_summary":"Diffusion-based re-ranking propagates similarity along a k-nearest-neighbour graph, but positive signal fades over several steps, so distant correct matches lose discriminative power. The paper proposes LPMT, which addresses this by encoding each instance as a probability distribution over its reciprocal local neighbourhood and then measuring the distance between two instances as the minimum cost of a multi-stage thermodynamic transition that moves one distribution to the other through intermediate local states. This transition cost is blended with the Euclidean feature distance to re-rank the initial retrieval results. Experiments on the revised Oxford and Paris landmark benchmarks (ROxf, RPar) and their 1M-distractor versions, using features from R-GeM, DOLG, CVNet, MAC, R-MAC, DELG, and SENet, report mAP improvements over diffusion, query expansion, and learning-based baselines. The motivation is that such a manifold-aware distance preserves local effectiveness while capturing global manifold structure, which is what the paper seeks to establish.","feed_headline":"Thermodynamic path distance sharpens image retrieval re-ranking","feed_subtitle":"Each image becomes a probability distribution; the minimum transition flow between them re-ranks image search results.","key_machinery":"The central object is the Thermodynamic Markovian Transition (TMT), whose defining identity is the equivalence between the minimum transition-flow cost of a master-equation Markov process and a path-sum of Wasserstein distances, $d'(i,j) = \\min_{\\pi} \\sum_k W_1(p_{i_k}, p_{i_{k+1}})$. That identity connects stochastic thermodynamics to optimal transport and makes the distance computable via entropy-regularized Sinkhorn iterations. The machinery also includes Bidirectional Collaborative Diffusion (BCD), which jointly optimizes aggregation weights and a bidirectional diffusion objective over an enlarged set of k-NN graphs to produce a robust similarity matrix, and Locality State Embedding (LSE), which converts each row of that matrix into a probability distribution over reciprocal neighbours. Together these components turn raw deep features into manifold-aware state distributions and a distance that can be used for re-ranking.","core_discovery":"The central claim is that a long-term thermodynamic Markovian transition over locally embedded probability distributions provides an effective manifold distance for instance retrieval. In the framework, each instance is represented by a sparse probability distribution built from the smoothed similarity matrix produced by Bidirectional Collaborative Diffusion and a k-reciprocal local region. The distance between two instances is then defined by the Thermodynamic Markovian Transition cost $d'(i,j) = \\min_{\\pi} \\sum_{k=0}^{K-1} W_1(p_{i_k}, p_{i_{k+1}})$, where $W_1$ is the earth-mover (Wasserstein) distance between consecutive state distributions; the paper proves in Appendix B.2 that this equals the minimum flow cost of a continuous-time Markov process governed by the master equation $\\dot{q}_t = T_t q_t$. The final retrieval distance is $d^*(i,j) = \\theta d(i,j) + (1-\\theta)d'(i,j)$, balancing Euclidean and manifold distances. The paper reports that this distance yields higher mAP than prior re-ranking methods across multiple feature extractors and benchmark protocols.","pith_inferences":["If the metric-collapse observation holds, the reported gains are likely attributable to BCD and LSE rather than to the multi-state thermodynamic stage; a locally graph-constrained transport cost would be a natural variant that could make the intermediate states genuinely change the optimum.","The distance formulation is feature-agnostic and could be applied to other retrieval settings, including text or multimodal embeddings, wherever a Euclidean baseline exists.","The entropy-regularized Sinkhorn solver introduces a temperature-like parameter; varying it could reveal whether the benefit of the transition cost comes from smoothing rather than from manifold geometry."],"forward_implications":["Re-ranking with LPMT improves mAP over query-expansion, diffusion-based, context-based, and learning-based baselines on ROxf, RPar, and their 1M-distractor versions across R-GeM, DOLG, CVNet, MAC, R-MAC, DELG, and SENet features.","Because LSE only uses reciprocal local neighbours, the distance is more robust to erroneous or missing k-NN graph connections than standard diffusion.","LPMT transfers to unsupervised content-based image retrieval, improving mAP and Recall@1 on CUB200, Indoor, and Caltech101.","Restricting re-ranking to the top-k images reduces the O(n^3) computation to O(k^3), keeping runtime under a few seconds for k=5000.","Ablations show the TMT distance beats cosine, Euclidean, Jaccard, and total-variation distances between the same embeddings, indicating the transition cost itself carries the improvement."],"supporting_citations":[{"why":"Supplies the master equation defining the thermodynamic transition flow.","marker":"(Seifert, 2012)"},{"why":"Underlies the claimed equivalence between transition flow cost and Wasserstein distance via local detailed balance.","marker":"(Van Vu & Saito, 2023)"},{"why":"Provides the entropy-regularized Sinkhorn algorithm used to compute the Wasserstein distances.","marker":"(Cuturi, 2013)"},{"why":"The Cluster-aware Similarity Diffusion objective that BCD extends with learnable graph weights.","marker":"(Luo et al., 2024)"},{"why":"Defines the bidirectional context diffusion objective that appears inside BCD's per-graph cost.","marker":"(Bai et al., 2019a)"},{"why":"The manifold-ranking diffusion formulation that motivates the k-NN graph construction.","marker":"(Zhou et al., 2003)"},{"why":"Efficient diffusion re-ranking baseline whose conjugate-gradient trick is reused in BCD.","marker":"(Iscen et al., 2017)"}],"fun_headline_variants":["Thermodynamic Markovian distance sharpens image retrieval","Locality preserving transition improves image retrieval","Manifold distance via Markovian transition improves retrieval","New transition cost for instance retrieval re-ranking","LPMT: thermodynamic transition for sharper image search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-step cost in the multi-state path is not a true metric, so the intermediate states can change the computed distance; the earth-mover distance is a metric, so as written the minimum over paths collapses to the direct earth-mover distance and the thermodynamic stages have no effect.","fun_headline_variants_meta":{"raw":{"variants":["Thermodynamic Markovian distance sharpens image retrieval","Locality preserving transition improves image retrieval","Manifold distance via Markovian transition improves retrieval","New transition cost for instance retrieval re-ranking","LPMT: thermodynamic transition for sharper image search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3737,"prompt_tokens":914,"completion_tokens":2823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2752}},"tokens_in":530,"tokens_out":2823,"duration_ms":25439,"temperature":1.0,"reasoning_tokens":2752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:24:16.372746+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $d'(i,j)$ from Eq. (20) with one intermediate state and with many for a sample of image pairs; since $W_1$ is a metric, the values coincide, which would show the multi-state transition contributes nothing and any gains come from LSE, BCD, or the Euclidean blend. Alternatively, check the released implementation to see whether $W_1$ is restricted to transitions along local graph edges; if it is not, the computed distance reduces to the direct earth-mover distance.","supporting_citations":[{"cited_title":"Stochastic thermodynamics, fluctuation theorems and molecular machines","cited_arxiv_id":null,"evidence_quote":"Supplies the master equation defining the thermodynamic transition flow."},{"cited_title":"and Saito, K","cited_arxiv_id":null,"evidence_quote":"Underlies the claimed equivalence between transition flow cost and Wasserstein distance via local detailed balance."},{"cited_title":"Sinkhorn distances: Lightspeed computation of optimal transport","cited_arxiv_id":null,"evidence_quote":"Provides the entropy-regularized Sinkhorn algorithm used to compute the Wasserstein distances."},{"cited_title":"Cluster-aware similarity diffusion for instance retrieval","cited_arxiv_id":null,"evidence_quote":"The Cluster-aware Similarity Diffusion objective that BCD extends with learnable graph weights."},{"cited_title":"Ranking on data manifolds","cited_arxiv_id":null,"evidence_quote":"The manifold-ranking diffusion formulation that motivates the k-NN graph construction."},{"cited_title":"Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations","cited_arxiv_id":null,"evidence_quote":"Efficient diffusion re-ranking baseline whose conjugate-gradient trick is reused in BCD."}],"review_version":1}