{"id":"032588c6-7d73-4308-aa1d-5cf4f365ada5","arxiv_id":"2507.19898","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A web-based visualization tool exposes Thompson Sampling's posterior draws, evidence counts, and selection history to support per-arm verification and step-level explanation.","lead":"TS-Insight is a new visual dashboard that reveals, arm by arm, how Thompson Sampling bandit algorithms update beliefs and pick actions over time. It is aimed at developers who need to debug, verify, and explain automated choices in high-stakes settings such as clinical trials and systematic literature review.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HDR Evolution Plot labels a draw outside its 50% symmetric interval as a 'rare event'; by construction 50% of draws fall outside, and the Appendix D interval is not a true HDR for skewed posteriors, so the tool's reliability signal is statistically unreliable.","rationale":"The reader's weakest assumption identifies the same component: the symmetric interval around the posterior mean in Appendix D is not a valid HDR for skewed Beta posteriors and can mislabel draws as rare. I agree that this is the most load-bearing weakness among the technical claims. My stress-test adds a second, independent failure: even when the bisection converges and the interval contains mass ρ exactly, ρ=0.50 means the 'rare event' label fires 50% of the time. This is a statistical category error, not just an approximation issue. It directly affects the paper's third key question and the claim that TS-Insight supports 'Understanding Sample Uncertainty and Risk.' The central contribution — a step-level visual audit of posterior draws, evidence counts, and selection histories — remains plausible and useful, so the concern does not warrant rejection. It does warrant a conditional acceptance: either compute a true HDR (or conventional central interval) at a genuinely rare threshold such as 95%, or soften the 'rare event' language. The paper itself acknowledges in Section 5 that formal user studies are future work, so the conditional stance is consistent with the presented evidence. My read therefore leaves the reader's CONDITIONAL verdict unchanged, but sharpens the technical reason for conditionality.","tokens_in":5443,"tokens_out":8191,"duration_ms":85800,"concrete_test":"Implement the exact HDR for the same DTS run used in Fig. 1 (or a synthetic Beta(0.5,100) arm), using a fine grid over θ and thresholding f(θ) ≥ λ to enclose probability mass ρ=0.50, then compare it with the tool's symmetric band from Appendix D. Count the fraction of posterior draws flagged 'rare' (outside the tool's band) across the trace. If the fraction is ≈50% (or if the exact HDR and tool band differ materially for skewed arms), the 'rare event' signal in Section 3 is misleading and the reliability claims require revision. Repeating the comparison with ρ=0.95 would show the intended rare-event semantics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's claim to support verification, diagnosis, and reliability rests on the HDR Evolution Plot (Section 3, Fig. 1A) and the 'rare event' signal it provides. Two distinct properties break this signal. First, Appendix D computes a symmetric interval around the posterior mean, [μ−δ, μ+δ], not a Highest Density Region. For any skewed Beta posterior (e.g., α=0.5, β=100), the density is monotone decreasing on [0,1]; the true ρ-level HDR is one-sided [0, q], while the symmetric interval includes a low-density right tail and excludes high-density values near 0, so the most probable draws are labeled rare. Second, the implementation targets ρ=0.50 (Appendix D: 'The resulting 50% HDR...'). A 50% band has exactly 50% of posterior draws outside it by construction, yet Section 3 tells users that a draw outside the band 'signals a rare event.' This makes the third key question ('When was the algorithm's choice outside the certainty region?') unreliable: developers will see a 'rare event' marker on half of all steps, not on genuinely unusual draws. The bisection in Eq. (1) can additionally fail to reach ρ when the symmetric interval's maximum mass (at δmax = min(μ, 1−μ)) is below ρ, silently returning a narrower band. These flaws do not invalidate the tool's forensic value for tracing α/β updates and selections, but they undermine the specific 'sample uncertainty and risk' analysis that is a stated contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TS-Insight, an open-source visual analytics tool for inspecting Thompson Sampling (TS) and Discounted Thompson Sampling (DTS) at the level of individual arms and steps. For each arm, the tool provides three synchronized subplots: an HDR Evolution Plot showing the posterior distribution and the location of posterior draws, an Alpha/Beta plot tracing the pseudo-counts of successes and failures, and a Barcode view summarizing the arm's selection history and outcomes. A separate XAI Snapshot view decomposes each selection into the posterior mean versus the posterior draw to explain why a particular arm was chosen. The paper claims these views support verification of the algorithm's mechanics, diagnosis of its behavior, and explanation of individual decisions. The contribution is illustrated with a walkthrough of a DTS-based active learning run on the SYNERGY systematic review dataset, and a preliminary validation on that dataset is mentioned but not reported. The paper is an accepted IEEE VIS 2025 poster, and its main text is accompanied by appendices containing additional references, a scope statement, key concept explanations, and a description of the HDR computation.","tokens_in":5748,"tokens_out":3983,"duration_ms":39854,"significance":"If the claims hold, TS-Insight addresses a real gap: existing visualizations of bandit algorithms typically show aggregate metrics such as cumulative regret, whereas developers debugging TS-based systems need step-level, per-arm traces of posterior updates and decisions. The paper explicitly provides machine-checkable code (the tool is open source), a concrete walkthrough, and a clear articulation of the tool's intended questions. The three-part decomposition of posterior belief, evidence, and outcome history is a sensible design, and the XAI Snapshot view's contrast between the posterior mean and the realized draw gives a concise, faithful explanation of why Thompson Sampling selected a particular arm. The major caveat is that the HDR Evolution Plot, which carries the paper's reliability/uncertainty analysis, does not in fact compute a Highest Density Region as claimed, and its 'rare event' signal is statistically unreliable as implemented. The paper also provides no empirical validation of the claimed verification, diagnosis, and explanatory benefits, beyond a single illustrative example.","major_comments":[{"comment":"The interval computed by the bisection in Appendix D is a symmetric interval around the posterior mean, not a Highest Density Region. For skewed Beta posteriors (e.g., α=0.5, β=100) the density is monotone decreasing, so the true HDR is one-sided, while the symmetric interval excludes high-density values near 0 and includes low-density values near 1. The density-ordering condition stated in Appendix D (f(x) ≥ f(y) for all x in the interval and y outside) is therefore violated for such posteriors. Consequently, the plot can label the most probable draws as lying outside the 'HDR' and low-probability draws as inside, inverting the intended 'rare event' signal that underlies the third key question in Section 4 ('When was the algorithm's choice outside the certainty region?'). The authors should either implement a true HDR computation (e.g., via the algorithm in Hyndman 1996) or rename the band as a 'central interval' and revise the interpretation accordingly.","section":"Appendix D, Eq. (1); Section 3, HDR Evolution Plot"},{"comment":"The implementation targets a 50% coverage (Appendix D: 'The resulting 50% HDR...'). By definition, 50% of posterior draws fall outside any 50% central interval, so the HDR Evolution Plot will mark roughly half of all draws as 'rare events' on every step. This makes the reliability signal uninformative at best and actively misleading at worst: a developer will see a 'rare event' marker at half the steps, not only for genuinely low-probability draws. In addition, the bisection loop in Appendix D can silently fail to reach the target probability ρ when the maximum enclosed mass, attained at δmax = min(μ, 1−μ), is less than ρ; the algorithm returns a narrower band than requested without warning. The authors should document the target coverage, justify the choice of 50%, and add a check for the case where the requested ρ is unattainable within the symmetric interval.","section":"Appendix D, bisection loop; Section 3, 'rare event' wording"},{"comment":"The abstract and introduction claim that TS-Insight 'enabl[es] verification, diagnosis, and explainability' of TS/DTS, but the paper provides no empirical evaluation: no user study, no quantitative task-based evaluation, and no results from the SYNERGY validation that is mentioned in Section 1. The only evidence is a single illustrative walkthrough (Fig. 1 and the surrounding text in Section 4). Given that the paper proposes a visual analytics tool with specific cognitive claims (verification, diagnosis, reliability assessment), the authors should either report the results of the mentioned SYNERGY validation or explicitly scope the paper as a design proposal and temper the abstract's language. At minimum, a task-oriented evaluation with a small number of developers would substantiate the central claim.","section":"Abstract and Section 1"}],"minor_comments":[{"comment":"The header contains a typo: 'T echnology' appears in the author affiliation lines; please correct this to 'Technology'.","section":"Title page"},{"comment":"The caption of Fig. 2 states the example is at 'Step t = 228' and the text describes Arm 7 being chosen with a posterior draw near 1.0, but the figure itself is not referenced in the body text with a specific analysis of all arms; consider adding a more detailed explanation of the log-scale y-axis and the relationship between the bars for μ and θ to help readers interpret the snapshot view.","section":"Figure 2"},{"comment":"The definitions of 'exploitation' and 'exploration' in Section 2 are outcome-based (whether the selected arm had the highest posterior mean), which conflates the realized event with the algorithm's latent random strategy. This is acceptable as a pedagogical simplification, but the text should note that the classification is an ex-post interpretation, not a property of the decision rule itself.","section":"Section 2"},{"comment":"The definition of Highest Density Region ('the range of values where we are 50% certain the arm's true success rate lies') is correct in spirit but would benefit from a mention that for skewed distributions the HDR need not be symmetric, anticipating the issue raised in Appendix D.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short poster-style contribution, and the central tracing/explanation design is useful and likely of interest to the VIS and bandit communities. However, the HDR computation error is a substantive correctness issue that directly affects the reliability/uncertainty analysis advertised as a key contribution. The fix is local—implement a true HDR or relabel the band—but it requires a revision of the text and possibly the figure. The lack of any evaluation is also a concern for a journal-level acceptance; if the revision adds a proper HDR and at least a modest evaluation of the tool's usability, the paper could become a minor-revision case. As it stands, I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TS-Insight is a genuinely useful idea: the first tool I know of that lets a developer watch Thompson Sampling step by step, with per-arm posterior draws, evidence counts, and selection history. The three-subplot decomposition and the XAI snapshot are well chosen, and the authors correctly summarize the TS/DTS mechanics. The code is on GitHub, which helps reproducibility. For a poster contribution, the paper is honest: it explicitly says no user study has been done and calls the SYNERGY usage a preliminary validation without showing results. I'd happily cite it as a starting point for TS visualization work.\n\nThe soft spot is not the lack of a study; it's the HDR Evolution Plot's statistical logic. The stress-test note is right. The paper computes a symmetric interval around the posterior mean, not a true Highest Density Region. For a Beta(0.5, 100), the density is monotone decreasing, so the real HDR is one-sided; a symmetric interval mislabels the most probable draws as rare. And because the target is 50%, exactly half the draws lie outside the band by construction. The text's phrase 'when a draw falls outside this region, it signals a rare event' is misleading — a 50% interval will be outside half the time. The bisection can also silently fail when the symmetric interval's max mass is below ρ. These flaws directly undercut the 'reliability' claim in Section 4. They don't break the barcode or alpha/beta tracing, which remain useful for verification and diagnosis.\n\nAlso minor: the exploration/exploitation label is a bit too simple — a draw above the mean isn't necessarily 'exploration' in a meaningful sense; it's just the sampling mechanism. That's a nitpick, though.\n\nIf this came to me for peer review, I would send it out, because the tool fills a real gap and the central design is sound. But I'd ask for a major revision before acceptance: either implement a true HDR (Hyndman's algorithm is a few lines) or stop calling it an HDR and frame the band explicitly as a symmetric credible interval, and soften or remove the 'rare event' language. With that, this is a solid tool note.","headline":"Useful TS visual debugging tool, but its 'rare event' signal is statistically broken; worth a revision.","tokens_in":6282,"tokens_out":3072,"would_cite":false,"duration_ms":29275,"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":"TS-Insight lets developers verify and explain Thompson Sampling decisions by turning each arm's posterior, evidence, and draw history into synchronized visual panels.","keywords":["Thompson Sampling","multi-armed bandits","visual analytics","explainable AI","algorithm visualization","Discounted Thompson Sampling","Highest Density Region","active learning"],"falsifier":"Take the posterior Beta(0.5,10) with target probability 0.5 and compute the symmetric interval that the paper's bisection returns around the mean 0.5/10.5; compute the probability mass actually inside that interval and compare the density at its endpoints with densities just outside it. If the mass is below 0.5, or if values outside the interval have higher density than values inside it, then the symmetric HDR is not a valid highest-density region and the rare-event flag is unreliable for skewed posteriors.","tokens_in":5199,"feed_emoji":"📊","tokens_out":12504,"duration_ms":108298,"temperature":0.7,"pith_summary":"Thompson Sampling and its variants make decisions by drawing a random sample from each arm's posterior belief and picking the arm with the highest draw, but their behavior is usually inspected only through aggregate metrics such as cumulative regret. TS-Insight is introduced as a visual analytics tool that makes these internal mechanics inspectable: for each arm it plots the evolving Beta posterior as a shaded Highest Density Region band, tracks the raw success and failure evidence counts, and records a barcode of when the arm was chosen and whether the pull succeeded. A separate XAI Snapshot view shows, at any step, every arm's posterior mean alongside its posterior draw, so a developer can see whether a selection was driven by strong belief or by an optimistic rare draw. The authors apply the tool to Discounted Thompson Sampling in an active-learning task on a fully labeled corpus of scientific papers, where the decay of the evidence counts on idle arms makes the discounting mechanism visible. The claim that grounds the paper is that with these views, a developer can verify the intended update and discounting logic, explain why a particular arm was selected, and judge when a choice falls outside the algorithm's own certainty region.","feed_headline":"TS-Insight shows Thompson Sampling at work, step by step","feed_subtitle":"Per-arm plots of beliefs, evidence, and draws turn exploration choices into inspectable events.","key_machinery":"The load-bearing object is the Highest Density Region (HDR) band drawn around each arm's posterior mean $\\mu_k = \\alpha_k/(\\alpha_k+\\beta_k)$. For a target probability $\\rho$, the tool approximates a symmetric interval $[a_\\rho,b_\\rho]$ by bisection on the Beta cumulative distribution function, clipping the endpoints to $[0,1]$ so that the interval is meant to contain probability mass $\\rho$, and shades that interval as the arm's certainty region. Each posterior draw $\\hat{\\theta}_k$ from the current Beta posterior is marked against the band, so a draw outside the band is read as a rare event signaling high epistemic uncertainty. The three subplots are synchronized: the HDR band is regenerated from the evidence counts, the barcode records which arm was selected and whether the reward was a success or failure, and the XAI Snapshot view compares $\\mu_k$ and $\\hat{\\theta}_k$ across all arms at one step to expose whether selection followed strong belief or a high but unlikely draw.","core_discovery":"The paper's central claim is that the opacity of Thompson Sampling can be removed by decomposing each arm's state into synchronized visual signals rather than by looking only at outcome-level aggregates. In TS-Insight, each arm has a row of three subplots: a Highest Density Region band around the posterior mean with each posterior draw marked against it, the cumulative success and failure counts, and a barcode of selection outcomes. At a selected step, the XAI Snapshot view juxtaposes every arm's posterior mean with its posterior draw, making it directly visible that the chosen arm is the one with the highest draw even when its mean is not the highest; in the running example at step 228, Arm 8 has the highest mean but Arm 7 is selected because its draw is higher, which the tool labels as exploration. The design is demonstrated on Discounted Thompson Sampling, where the decay of the evidence counts on idle arms becomes visible in the evidence subplot, allowing a developer to trace a single decision back to the belief and evidence that produced it. In the authors' reading, this is what makes verification, diagnosis, and explainability of exploration/exploitation dynamics possible.","pith_inferences":["A direct test of the tool's value would be a controlled debugging study with injected faults, such as a wrong discount factor, a reversed reward mapping, or a skipped posterior update; the paper lists user studies as future work but does not propose this specific fault-injection protocol.","The symmetric HDR approximation should be stress-tested on strongly skewed Beta posteriors, for example $\\alpha=0.5, \\beta=10$ with $\\rho=0.5$; because the true highest-density region is one-sided there, the symmetric band can misclassify ordinary draws as rare, which is an implication of the design rather than a claim the paper makes.","Since the visual machinery only needs a posterior, an evidence count, and a sampled outcome per arm, the same decomposition could be applied to other Thompson Sampling variants such as Sliding-Window or Batched TS, which the paper names but does not visualize.","The snapshot view suggests a cheap diagnostic not stated in the paper: if many selections are attributed to draws far outside the HDR band, the algorithm may be exploring more aggressively than its posterior uncertainty justifies, and TS-Insight could aggregate this into a summary statistic."],"forward_implications":["A developer can verify the update logic by locating a pull in the barcode and seeing the corresponding change in the success or failure count and the resulting narrowing or widening of the HDR band on that arm.","In Discounted Thompson Sampling, the forgetting mechanism becomes visible as idle arms' evidence counts decay toward zero, letting users confirm that recent evidence is being prioritized as intended.","At every step, the XAI Snapshot view distinguishes exploration from exploitation: if the chosen arm is not the one with the highest posterior mean, the tool attributes the choice to a higher posterior draw.","A draw falling outside the HDR band gives a per-arm reliability signal, warning that the algorithm acted on a low-probability outcome while the posterior was uncertain or unstable.","Hiding or showing arms and step ranges lets a user compare arm trajectories across phases of the run, which can reveal non-stationarity in the environment as shifts in which arms are selected."],"supporting_citations":[{"why":"introduces the original Thompson Sampling rule that the tool visualizes, selecting an arm by the posterior probability that it is best.","marker":"[6]"},{"why":"supplies the modern formulation of Thompson Sampling as a Bayesian heuristic and the stationary-environment assumption behind the posterior evolution plots.","marker":"[5]"},{"why":"defines Discounted Thompson Sampling, the variant whose discounting mechanism is the paper's main verification target.","marker":"[4]"},{"why":"provides the method for computing and graphing Highest Density Regions that the belief subplot is built on.","marker":"[3]"},{"why":"motivates the application scenario of algorithm selection for deep active learning, which the tool is used to debug.","marker":"[7]"},{"why":"supplies the fully labeled scientific-paper corpus used in the paper's preliminary validation of TS-Insight.","marker":"[1]"}],"fun_headline_variants":["TS-Insight: See why Thompson Sampling picks an arm","Visualize Thompson Sampling's beliefs and draws per arm","TS-Insight turns Thompson Sampling black box transparent","Trace Thompson Sampling decisions to posterior draws","TS-Insight: Inspect each arm's evidence and selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The tool's ability to flag a posterior draw as a rare event assumes that the symmetric interval it draws around each arm's average belief is the region where the belief's density is actually highest; for strongly skewed Beta posteriors that interval is not the highest-density region, so an ordinary draw can be mislabeled as rare, and when the average sits at 0 or 1 the code erases the band entirely.","fun_headline_variants_meta":{"raw":{"variants":["TS-Insight: See why Thompson Sampling picks an arm","Visualize Thompson Sampling's beliefs and draws per arm","TS-Insight turns Thompson Sampling black box transparent","Trace Thompson Sampling decisions to posterior draws","TS-Insight: Inspect each arm's evidence and selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1408,"prompt_tokens":889,"completion_tokens":519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":441}},"tokens_in":505,"tokens_out":519,"duration_ms":5297,"temperature":1.0,"reasoning_tokens":441,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:51:46.861930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the posterior Beta(0.5,10) with target probability 0.5 and compute the symmetric interval that the paper's bisection returns around the mean 0.5/10.5; compute the probability mass actually inside that interval and compare the density at its endpoints with densities just outside it. If the mass is below 0.5, or if values outside the interval have higher density than values inside it, then the symmetric HDR is not a valid highest-density region and the rare-event flag is unreliable for skewed posteriors.","supporting_citations":[],"review_version":1}