{"id":"35936825-1908-457f-a555-2b33be0c7438","arxiv_id":"2412.07299","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Bi-Mamba, a bidirectional state-space deep-learning model, infers diffusion coefficients and anomalous exponents from short trajectories with mid-tier AnDi-2 challenge results.","lead":"This paper tests Bi-Mamba, a bidirectional deep-learning model based on the Mamba state-space architecture, for reading diffusion behavior out of short, noisy particle trajectories. The model placed mid-tier in the AnDi-2 challenge, with a 7th place for the anomalous exponent and a 9th for the diffusion coefficient.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that Bi-Mamba scores 'better overall and in each category' rests on single-run loss curves without error bars, confidence intervals, or significance tests; this is the most load-bearing gap in the empirical comparison.","rationale":"The reader's weakest_assumption focuses on external validity, namely that AnDi-2 synthetic data faithfully represent experimental conditions. That concern is real but is about whether benchmark scores transfer to practice; it is harder to settle with a single decisive check. My concern targets the immediate internal support for the claim that Bi-Mamba is better than the bidirectional RNN baseline: the reported evidence is a single set of loss curves with no statistical quantification. This is the most directly falsifiable load-bearing assumption, and it can be settled by rerunning with multiple seeds. The reader does mention missing error bars and significance tests in the rationale, so there is partial agreement, but the stated weakest assumption is different. The paper deserves credit for providing a public GitHub link, which would make the proposed rerun feasible. The conditional verdict remains appropriate: acceptance should require the statistical validation described, not an outright rejection.","tokens_in":5918,"tokens_out":6907,"duration_ms":75342,"concrete_test":"Run both Bi-Mamba and the Bi-RNN baseline on the same AnDi-2 tasks using identical train/test splits, preprocessing, and hyperparameters for at least 10 independent seeds. Report mean plus/minus standard deviation and a paired significance test (e.g., Wilcoxon signed-rank or paired bootstrap) for MAE(alpha), MSLE(K), F1(diffusion type), and RMSE(change-point). If the per-category improvement over the RNN is not consistently positive and significant across seeds, the claim should be weakened from 'better' to 'comparable.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that Bi-Mamba scores better overall and in each category of the AnDi2 Challenge than a bidirectional RNN baseline, and is therefore promising for anomalous-diffusion inference. The evidence for this superiority claim is Fig. 3, which shows MAE, MSLE, and WCE loss curves on a test set of 10^4 trajectories. The results paragraph does not report how many independent training runs were performed, any error bars, confidence intervals, or paired significance tests; it also does not give the RNN baseline's quantitative AnDi2 scores or the hyperparameters used. Since deep-network training is stochastic, the 'better overall and in each category' assertion is not yet supported beyond a single run. The listed AnDi2 ranks (7th in alpha, 9th in K, 3rd in diffusion type, 10th in change-point) are point estimates without score distributions or participant counts, so even the weaker claim of competitiveness is hard to evaluate. If the reported advantage over the baseline is within run-to-run noise, the paper's main positive result disappears. This is load-bearing because the abstract's 'infers efficiently' and the conclusion's 'demonstrated a notable efficiency' rest on this comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Bi-Mamba, a bidirectional state-space deep-learning architecture, for characterizing anomalous diffusion from single, short trajectories. The model uses a three-block design that first segments the trajectory into diffusion states and then performs separate regressions of the effective diffusion coefficient K and the anomalous exponent alpha. The authors train and evaluate on the AnDi-2 challenge datasets, reporting loss curves against a bidirectional RNN baseline and listing their AnDi-2 challenge ranks. The central claim is that Bi-Mamba infers K and alpha efficiently and scores better overall and in each category than the RNN baseline.","tokens_in":6139,"tokens_out":4682,"duration_ms":46051,"significance":"If the comparative claim is substantiated, Bi-Mamba would be a useful addition to the toolbox for single-particle-tracking analysis, with the practical advantage of efficient, bidirectional sequence processing. Strengths of the work include public availability of the implementation on GitHub, use of the standardized AnDi-2 benchmark, and the inclusion of a bidirectional RNN as an internal baseline. However, the current evidence is not yet sufficient to support the headline superiority claim: the comparison rests on qualitative loss curves from what appears to be a single run, and the AnDi-2 ranks are reported without statistical context. The work is therefore promising but needs a substantially strengthened empirical evaluation.","major_comments":[{"comment":"The central claim that Bi-Mamba \"scor[es] better overall and in each category\" than the bidirectional RNN baseline is supported only by qualitative loss curves plotted from what appears to be a single training run. No error bars, confidence intervals, repeated-seed statistics, or paired significance tests are reported, so the reported advantage may lie within run-to-run variability. Because the abstract's \"infers efficiently\" and the conclusion's \"notable efficiency\" rest on this comparison, the authors should rerun the experiment with multiple seeds and report a quantitative comparison table with means and dispersions.","section":"Results paragraph and Fig. 3"},{"comment":"The listed AnDi-2 results (7th in alpha, 9th in K, 3rd in diffusion type, 10th in change-point) are point estimates without score distributions, participant counts, or the score gaps to neighboring teams. Without this context, a reader cannot tell whether these ranks indicate a meaningful improvement or a knife-edge placement, and the claim of competitiveness is therefore undersupported.","section":"AnDi-2 ranks paragraph"},{"comment":"The manuscript does not report the hyperparameters of Bi-Mamba or of the bidirectional RNN baseline (number of layers, state dimension, learning rate, batch size, epochs, optimization details), despite the text asserting a \"fair benchmarking\" comparison. Reproducibility requires these details, and the GitHub link alone does not substitute for reporting them in the paper.","section":"Architecture description and Fig. 2"},{"comment":"The total loss is described as the sum of the MAE, MSLE, and WCE losses, but no weighting or balancing scheme is given. Since these losses have different scales and units, an unweighted sum can be dominated by one task; the paper should specify the weighting and, ideally, show sensitivity of the results to it.","section":"Multitask loss paragraph"}],"minor_comments":[{"comment":"The text says diffusion modes are \"hot-encoded\"; this should read \"one-hot encoded\".","section":"Architecture paragraph (p. 3)"},{"comment":"There is a typo in \"Mean-Squared Logarithic Error\"; it should be \"Mean-Squared Logarithmic Error\".","section":"Evaluation paragraph (p. 3)"},{"comment":"The figure legend uses \"RNN\" and \"Mamba\", while the text compares a bidirectional RNN with Bi-Mamba; the caption should define both labels.","section":"Fig. 3 caption"},{"comment":"The trajectory count \"10 4\" in the Fig. 3 caption should be typeset as \"10^4\".","section":"Data description paragraph (p. 3)"},{"comment":"The phrase \"tested on the AnDi-2 challenge datasets among others\" is vague; the paper should either identify the \"other\" datasets or remove the phrase.","section":"Abstract and data availability"},{"comment":"Reference [15] is the AnDi-2 registered-report protocol; if the challenge results have been published, that version should be cited instead of or in addition to the protocol.","section":"Reference [15]"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is more of a short benchmark report than a full research article. The core architecture is sensible and the public code is a plus, but the empirical evidence for the central superiority claim is currently too thin. I would encourage the authors to provide repeated-run statistics, a quantitative comparison table, and full hyperparameter details; then the paper could be a defensible contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the application of a bidirectional Mamba state-space architecture to anomalous diffusion inference from short, noisy trajectories. That specific transfer looks like a first, and the paper does something right: it evaluates on the external AnDi-2 benchmark, reports its official ranks, and provides a code link. That is real evidence and deserves credit.\n\nThe architecture is described clearly enough: a segmentation block followed by dedicated K and alpha regression blocks, with a bidirectional scan borrowed from Bi-Mamba+. The task-specific blocking is sensible, and the loss curves in Fig. 3 show that Mamba trains more stably than the bidirectional RNN baseline (lower variance, no overfitting up to 500 epochs). That stability result is probably the most useful observation in the paper.\n\nThe soft spots are real and load-bearing. The claim that Bi-Mamba scores \"better overall and in each category\" of the AnDi2 Challenge rests entirely on Fig. 3, which shows loss curves without error bars, confidence intervals, or any statement about the number of independent runs. The \"smaller loss variance\" is asserted but not quantified. The official ranks (7th alpha, 9th K, 3rd diffusion type, 10th change-point) are point estimates with no score distributions or participant counts, so we cannot tell if those ranks are close or far from the leaders. Hyperparameters (layers, state dimension, learning rate, epochs) are not reported, which makes the baseline comparison hard to reproduce. None of these are fatal; they are fixable with a few independent training runs and a small table.\n\nThe math and data handling look fine. The MSD model, the five bio-mimetic AnDi-2 models, and the feature set (displacements, 1D MSDs, angle, distance from origin) are standard. No circularity: training on AnDi-2 train and evaluating on the held-out test set is the right protocol.\n\nWho gets value from this: anyone working on single-particle tracking inference who wants to know whether Mamba-style architectures are worth trying, and researchers thinking about transferring modern sequence models into physics. It is not a SOTA paper, but it is a legitimate incremental contribution on a timely topic. I would send it to peer review, with the request that the authors add error bars, hyperparameters, and a concrete comparison table to other AnDi-2 participants before acceptance. My own verdict is conditional, leaning positive.","headline":"Useful first Mamba-for-physics demo with honest AnDi-2 participation, but the headline comparison to the RNN baseline needs error bars before I'd trust it.","tokens_in":6702,"tokens_out":1926,"would_cite":false,"duration_ms":22474,"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":"Bi-Mamba, a bidirectional state-space deep-learning model, recovers the effective diffusion coefficient and anomalous exponent from single short noisy trajectories, outperforming a bidirectional RNN baseline on the AnDi-2 challenge.","keywords":["Anomalous diffusion","Single-particle tracking","State-space models","Mamba","Bidirectional scan","Deep learning","Diffusion coefficient","Anomalous exponent"],"falsifier":"Apply the trained Bi-Mamba model to experimental single-particle trajectories whose ground-truth $K$ and $\\alpha$ are known from an independent measurement — for example, a calibrated optical trap in a well-characterized viscoelastic fluid — and compare inferred values with that ground truth; errors substantially larger than the benchmark's reported errors would falsify the transfer claim.","tokens_in":5742,"feed_emoji":"🔬","tokens_out":8968,"duration_ms":84571,"temperature":0.7,"pith_summary":"Characterizing anomalous diffusion often comes down to estimating the effective diffusion coefficient $K$ and the anomalous exponent $\\alpha$ from particle trajectories that are short, sparse, and noisy. This paper argues that Bi-Mamba, a deep state-space architecture with a bidirectional scan mechanism, can infer both parameters from a single two-dimensional trajectory while also segmenting the trajectory into different diffusion states. The model is evaluated on the AnDi-2 challenge datasets, which contain five bio-mimetic models of single-particle motion, and the authors report that it scores better overall and in every category than a bidirectional RNN baseline. If that benchmark result carries over to real experiments, biophysicists would gain a practical tool for single-particle tracking in environments where only a few short tracks are available.","feed_headline":"Bi-Mamba extracts diffusion parameters from single short tracks","feed_subtitle":"A bidirectional state-space model scores above a recurrent baseline on every AnDi-2 challenge task.","key_machinery":"The central object is the Bi-Mamba block, defined by a Mamba layer — a selective state-space model that compresses a sequence into a hidden state while deciding, in an input-dependent way, what to remember and what to ignore — applied once to the trajectory features and once to their time-flipped counterparts, then concatenating the two outputs through a feedforward layer. The full model stacks three such blocks: one performs hot-encoded diffusion-state segmentation, one regresses $K$ under a mean-squared-logarithmic-error loss, and one regresses $\\alpha$ under a mean-absolute-error loss. The features fed in are the per-axis displacements, the one-dimensional mean-squared displacements, the angle between consecutive displacements, and the distance from the origin. This bidirectional design is what allows the model to capture time dependencies across the whole trajectory while keeping training linear in sequence length.","core_discovery":"On the paper's own terms, the central claim is that Bi-Mamba — a selective state-space model equipped with a bidirectional scan — can take one short, noisy, two-dimensional trajectory and reliably infer the effective diffusion coefficient $K$ and anomalous exponent $\\alpha$ (with $\\mathrm{MSD}(t) = 4 K t^\\alpha$), as well as identify changes in the diffusion state along the trajectory. Using the AnDi-2 challenge datasets, which represent five bio-mimetic models including single-state, multi-state, dimerization, transient confinement, and quenched-trap motion, the authors report that Bi-Mamba outperforms a bidirectional recurrent neural network baseline both overall and in each task category. They further report that the model trains for more epochs without overfitting and with lower loss variance, and they describe a three-block architecture in which segmentation, $K$ regression, and $\\alpha$ regression are handled by dedicated blocks with a summed loss.","pith_inferences":["The input-dependent selection mechanism may make Bi-Mamba robust to trajectories with more than two diffusion states, but the paper only tests the two-state scenarios of the challenge; this remains an open extension.","The hidden state at each time step could serve as a learned low-dimensional representation of trajectory dynamics, enabling clustering or anomaly detection beyond the two regression targets; the paper does not explore this.","The comparison against a single bidirectional RNN baseline does not by itself locate the source of the gain; a systematic comparison with Transformers or a unidirectional Mamba would separate the contribution of the bidirectional scan from the choice of state-space backbone."],"forward_implications":["Single-particle-tracking experiments with short, noisy trajectories could estimate $K$ and $\\alpha$ from individual tracks without averaging over many trajectories.","The same model can return both global parameters and a per-time segmentation, so binding and unbinding events can be read directly from one network output.","Because the model trains stably for a larger number of epochs, fine-tuning specialized versions on simulations generated near the general model's predictions is a practical route to further accuracy.","The authors' stated expectation that Mamba-based U-Net variants will improve segmentation points to a concrete next step for the architecture."],"supporting_citations":[{"why":"It defines the AnDi challenge benchmark and the bio-mimetic trajectory models used for training and evaluation.","marker":"[14]"},{"why":"It sets out the AnDi-2 tasks, including segmentation and inference of $K$ and $\\alpha$, whose scores the paper reports.","marker":"[15]"},{"why":"It introduces the Mamba selective state-space architecture that Bi-Mamba adapts.","marker":"[32]"},{"why":"It supplies the bidirectional-scan construction for Mamba that the paper builds upon.","marker":"[33]"},{"why":"It provides the displacement-angle feature used as one of the input descriptors.","marker":"[34]"},{"why":"It frames deep learning as a tool for advanced microscopy and single-particle tracking, which motivates the approach.","marker":"[13]"}],"fun_headline_variants":["Bidirectional Mamba outshines RNN on diffusion inference","State-space model extracts diffusion from short trajectories","Bi-Mamba excels at anomalous diffusion characterization","Short tracks, precise diffusion: Bi-Mamba leads","Mamba model beats baseline on every AnDi-2 task"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The AnDi-2 challenge's synthetic datasets, with their five bio-mimetic models, noise levels, and short trajectories, are representative enough of real experimental single-particle tracking data that the reported benchmark gains transfer to practical use.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional Mamba outshines RNN on diffusion inference","State-space model extracts diffusion from short trajectories","Bi-Mamba excels at anomalous diffusion characterization","Short tracks, precise diffusion: Bi-Mamba leads","Mamba model beats baseline on every AnDi-2 task"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2586,"prompt_tokens":816,"completion_tokens":1770,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1694}},"tokens_in":432,"tokens_out":1770,"duration_ms":16551,"temperature":1.0,"reasoning_tokens":1694,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:55:32.429299+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the trained Bi-Mamba model to experimental single-particle trajectories whose ground-truth $K$ and $\\alpha$ are known from an independent measurement — for example, a calibrated optical trap in a well-characterized viscoelastic fluid — and compare inferred values with that ground truth; errors substantially larger than the benchmark's reported errors would falsify the transfer claim.","supporting_citations":[{"cited_title":"Mu˜ noz-Gil, G","cited_arxiv_id":null,"evidence_quote":"It defines the AnDi challenge benchmark and the bio-mimetic trajectory models used for training and evaluation."},{"cited_title":"Mu˜ noz-Gil, H","cited_arxiv_id":null,"evidence_quote":"It sets out the AnDi-2 tasks, including segmentation and inference of $K$ and $\\alpha$, whose scores the paper reports."},{"cited_title":"Kabbech and I","cited_arxiv_id":null,"evidence_quote":"It provides the displacement-angle feature used as one of the input descriptors."},{"cited_title":"Midtvedt, S","cited_arxiv_id":null,"evidence_quote":"It frames deep learning as a tool for advanced microscopy and single-particle tracking, which motivates the approach."}],"review_version":1}