{"id":"89e9ac6e-cf0d-475d-90a9-7344693f41dc","arxiv_id":"2606.01325","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SEArch applies optimistic FTRL with adaptive learning rate under the Stochastically Extended Adversary model to achieve regret O(σ̄_T √T + √J) for multi-policy UAV radar search, with a windowed variant and up to 30% empirical improvement.","lead":"The paper introduces SEArch, an optimistic Follow-the-Regularized-Leader algorithm for selecting among radar detectors on UAVs to minimize regret under both measurement noise and scene transitions. A smart generalist might read it to see how online learning enables real-time adaptation for autonomous sensing platforms operating in changing conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Regret analysis requires real-time observability of per-detector losses, which is not guaranteed in UAV radar search without ground truth","rationale":"The reader's weakest_assumption already isolates the observability requirement for the library; the full abstract text confirms this is exactly the condition needed for the online selector to function. No other derivation step (e.g., the specific form of the adaptive learning rate or the windowed variant) appears internally inconsistent from the given material.","tokens_in":1870,"tokens_out":351,"duration_ms":17746,"concrete_test":"Re-derive the regret bound of SEArch under bandit feedback (only loss of chosen detector observed) instead of full information; if the bound degrades to O(T^{2/3}) or worse, the full-info assumption is load-bearing for the stated rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The SEA framework and SEArch (optimistic FTRL with adaptive rate) achieve the claimed O(σ̄_T √T + √J) bound by treating intra-scene noise as stochastic and inter-scene shifts as adversarial, with the selector updating on observed losses each round. This presupposes that, after selecting a detector from the fixed library, its loss (detection performance) is revealed or estimable onboard in real time. In the target UAV radar application (e.g., respiration micro-motion through occlusions), no such immediate ground-truth feedback exists; the UAV cannot independently verify whether a chosen detector correctly identified a target in the current scene. If only bandit or no feedback is available, the OFTRL update and the derived regret no longer apply, breaking the central guarantee that adaptation works without oracle scene knowledge.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper formulates UAV radar target search as an online policy selection problem over a fixed library of detectors. It introduces the Stochastically Extended Adversary (SEA) framework to jointly handle intra-scene stochastic noise and inter-scene adversarial shifts without oracle knowledge of dynamics, instantiates it via the lightweight optimistic FTRL algorithm SEArch (with adaptive rate) that achieves regret O(¯σ_T √T + √J), and proposes the windowed W-SEArch variant achieving O(¯σ_I √w) under at most one transition per window. Experiments are reported to yield up to 30% regret reduction versus non-adaptive baselines.","tokens_in":2041,"tokens_out":532,"duration_ms":22202,"significance":"If the regret analysis holds under the paper's feedback model, the SEA framework's explicit treatment of mixed stochastic-adversarial non-stationarity, together with the resource-light online selector, would constitute a useful contribution to adaptive perception on constrained platforms. The explicit regret decomposition separating measurement noise from scene transitions is a clear strength, as is the claim of no oracle scene dynamics. The reported experimental gains, if supported by full details, would strengthen the practical case.","major_comments":[{"comment":"Abstract: the SEA framework and SEArch regret bound O(¯σ_T √T + √J) are derived under the assumption that the loss of the selected detector is observed (or estimable) after each round to enable the OFTRL update. The target UAV radar application (respiration micro-motion through occlusions) provides no such immediate ground-truth feedback, so the update rule and the derived bound do not apply as stated; this assumption is load-bearing for the central claim that adaptation works without oracle knowledge.","section":"Abstract"},{"comment":"Abstract: the phrasing 'without requiring oracle knowledge of scene dynamics' addresses only the inter-scene transitions J but leaves unaddressed the per-round loss observability required by the optimistic FTRL selector; the two are distinct and the latter is necessary for the regret guarantee to transfer to the UAV setting.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract states 'Experiments show up to 30% regret reduction' but supplies no information on the number of trials, scene-transition models, detector library size, or how losses were computed in simulation; these details are needed for reproducibility even if the main text contains them.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful reading and for highlighting the feedback model assumptions. We agree that the per-round loss observability required by the OFTRL update is a distinct and load-bearing assumption that was not sufficiently distinguished from the lack of oracle scene dynamics in the abstract and introduction. We will revise the manuscript to clarify this point explicitly.","responses":[{"response":"We acknowledge that the regret analysis is derived under the standard full-information feedback model of online learning, in which the loss of the played action is observed after each round. The UAV radar application indeed lacks immediate ground-truth labels. We will revise the abstract, introduction, and a new dedicated paragraph in Section 2 to state the feedback assumption explicitly, to separate it from the 'no oracle scene dynamics' claim, and to discuss practical loss estimation via proxy radar statistics or delayed feedback. The SEA framework itself remains valid under the stated feedback model; the revision will make the scope of the guarantees transparent.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the SEA framework and SEArch regret bound O(¯σ_T √T + √J) are derived under the assumption that the loss of the selected detector is observed (or estimable) after each round to enable the OFTRL update. The target UAV radar application (respiration micro-motion through occlusions) provides no such immediate ground-truth feedback, so the update rule and the derived bound do not apply as stated; this assumption is load-bearing for the central claim that adaptation works without oracle knowledge."},{"response":"We agree that the current phrasing conflates the two issues. The sentence will be rewritten to read: 'without requiring oracle knowledge of scene transition times or dynamics, under the assumption that per-round losses are observed or estimable.' Corresponding clarifications will appear in the model section and the abstract.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the phrasing 'without requiring oracle knowledge of scene dynamics' addresses only the inter-scene transitions J but leaves unaddressed the per-round loss observability required by the optimistic FTRL selector; the two are distinct and the latter is necessary for the regret guarantee to transfer to the UAV setting."}],"tokens_in":1586,"tokens_out":483,"duration_ms":21982,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main contribution is the SEA model that treats intra-scene radar noise as stochastic and inter-scene changes as adversarial, then derives an OFTRL selector (SEArch) whose regret scales as O(σ̄_T √T + √J) without needing to know the transition times in advance. W-SEArch adds a windowed restart for faster adaptation. This is a modest but concrete step beyond standard online learning that directly targets the UAV radar use case.\n\nThe work is clear on the modeling choice and the resulting bounds. The 30% regret reduction in the reported experiments is the sort of number that makes the idea worth checking in a real system.\n\nThe soft spot is the feedback assumption. The regret derivation relies on observing the loss of the chosen detector each round so the optimistic update can run. In the respiration-micro-motion search scenario, the UAV has no immediate ground truth to score how well a detector performed on a given scene; any loss estimate would have to come from indirect or delayed signals. If the paper only assumes full information feedback, the central guarantee does not transfer to the stated application. The abstract does not resolve this gap.\n\nThis is for researchers who already work on online policy selection for resource-constrained sensing. It is worth sending to review so the authors can show either how losses are estimated in practice or how the bounds change under bandit feedback. I would not cite it until that point is settled.","headline":"SEArch gives a clean extension of optimistic FTRL to mixed stochastic-adversarial regret for detector switching, but the UAV radar setting may not supply the per-round loss observations the analysis needs.","tokens_in":2546,"tokens_out":376,"would_cite":false,"duration_ms":13489,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"An online selector over a library of radar detectors achieves regret that scales with noise level times square root of time plus square root of scene transitions, without oracle knowledge of when scenes change.","keywords":["UAV radar search","online policy selection","regret bounds","non-stationary environments","radar detectors","adaptive learning rate","SEA framework"],"falsifier":"Collect regret traces from SEArch and W-SEArch on a sequence of radar scenes whose transition count J and per-scene noise levels are known in advance, then check whether the measured cumulative regret stays inside the envelope O(σ̄_T √T + √J) or deviates systematically.","tokens_in":2757,"feed_emoji":"🛸","tokens_out":717,"duration_ms":21573,"temperature":0.7,"pith_summary":"The paper casts UAV radar target search as an online decision problem in which the vehicle must pick which specialized detector to run at each step, because radar statistics shift as the platform moves through different scenes. It models the combined effect of within-scene measurement noise and between-scene shifts with the Stochastically Extended Adversary framework and then supplies a lightweight optimistic Follow-the-Regularized-Leader algorithm, SEArch, whose regret is bounded by the product of average noise and the square root of mission length plus the square root of the number of transitions. A windowed variant, W-SEArch, restarts periodically to keep regret small even when transitions occur frequently.","feed_headline":"Online selector adapts UAV radar detectors across shifting scenes","feed_subtitle":"SEArch bounds regret by noise times square root of time plus square root of transitions without needing to know when scenes change.","key_machinery":"The Stochastically Extended Adversary (SEA) model that couples stochastic noise inside each scene with adversarial shifts across scenes; SEArch performs policy selection with optimistic Follow-the-Regularized-Leader using an adaptive learning rate.","core_discovery":"SEArch instantiates the SEA framework via an optimistic Follow-the-Regularized-Leader selector equipped with an adaptive learning rate and obtains regret O(σ̄_T √T + √J), where σ̄_T captures radar measurement noise and J counts scene transitions over horizon T; the windowed W-SEArch variant that restarts every w rounds obtains O(σ̄_I √w) regret provided there is at most one transition per window.","pith_inferences":["The same selector structure could be applied to other onboard perception tasks where several algorithms compete under shifting conditions, such as camera-based object detection.","Real-time estimation of local noise level σ̄ becomes the practical bottleneck once the theoretical bound is accepted.","Dynamically choosing window size w according to recently observed transition frequency could tighten the regret further."],"forward_implications":["Regret grows only with the square root of the number of scene changes rather than linearly, so longer missions remain feasible.","W-SEArch keeps regret controlled even under rapid scene changes by resetting every w steps.","No external forecast of scene boundaries is required; the algorithm adapts using only observed losses.","Experiments report up to 30 percent lower regret than non-adaptive baselines across varied non-stationary radar settings."],"fun_headline_variants":["SEArch adapts UAV radar policies across noise and scene drifts","W-SEArch restarts windows to bound regret under one transition","OFTRL selector achieves UAV radar regret bound without scene oracle","SEA accounts for stochastic noise and inter-scene shifts in UAV search"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"A fixed library of specialized detectors exists whose per-scene performance can be observed or estimated in real time by the resource-limited UAV.","fun_headline_variants_meta":{"raw":{"variants":["SEArch adapts UAV radar policies across noise and scene drifts","W-SEArch restarts windows to bound regret under one transition","OFTRL selector achieves UAV radar regret bound without scene oracle","SEA accounts for stochastic noise and inter-scene shifts in UAV search"]},"model":"grok-4.3","cost_usd":0.005381,"raw_usage":{"total_tokens":3909,"prompt_tokens":2868,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":53811000,"prompt_tokens_details":{"text_tokens":2868,"audio_tokens":0,"image_tokens":0,"cached_tokens":768},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":973,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":2868,"tokens_out":68,"duration_ms":38911,"temperature":1.0,"reasoning_tokens":973,"cache_read_input_tokens":768,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T16:08:53.527466+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Collect regret traces from SEArch and W-SEArch on a sequence of radar scenes whose transition count J and per-scene noise levels are known in advance, then check whether the measured cumulative regret stays inside the envelope O(σ̄_T √T + √J) or deviates systematically.","supporting_citations":[],"review_version":1}