{"id":"065fac3c-2f76-49f7-892e-2aaf500f9be7","arxiv_id":"2607.10990","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Dynamic DFS tree reasoning with adaptive tool routing yields SOTA SoccerBench scores (85.2/87.4/82.2%) and 74.16% on NExT-QA.","lead":"TreeSoc answers hard soccer video questions by growing a search tree of sub-tasks and calling specialized vision and knowledge tools at each node. It reports large gains on SoccerBench and solid zero-shot transfer to general video QA, arguing that structured tool use beats single-pass VLMs for multi-step sports reasoning.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"SOTA and paradigm claims rest on tool quality more than proven DFS/replanning; no ablation isolates the adaptive tree mechanism.","rationale":"The reader correctly flags reliability of intermediate tools and Φ_LLM reflection as the weakest assumption and notes missing causal ablations of DFS/replanning. That is exactly the load-bearing gap: Table 1 SOTA numbers and the paradigm language in Abstract/§6 are presented as evidence for hierarchical search, but the experimental design never holds tools fixed while ablating the dynamic tree. SoccerAgent is the closest agentic baseline, yet tool sets and routing differ, so the comparison cannot isolate DFS. NExT-QA success without soccer tools further suggests general LLM+decomposition skill rather than the claimed soccer tree synergy. A same-tool static-plan control would settle whether the adaptive mechanism is necessary. Until then CONDITIONAL remains appropriate: engineering contribution is real if code works, but the “tree reasoning as effective paradigm” claim is not yet causally supported. No stronger rejection is warranted; the paper is clear and conventional for systems work.","tokens_in":11503,"tokens_out":587,"duration_ms":6112,"concrete_test":"On the SoccerBench public test, re-run TreeSoc with identical tools and LLM but force a static plan: disable Reflect/Update-Queue (always Keep, no Insert/Skip/Retry) and use a single non-recursive decomposition at depth 1. If Text/Image/Video accuracies drop by less than ~3–5 points relative to full DFS, the adaptive-tree contribution is not load-bearing for the SOTA claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Abstract; §5.2; Table 1) attributes SOTA SoccerBench scores and the “effective paradigm” of structured tree reasoning primarily to dynamic DFS decomposition plus Reflect/Update-Queue replanning (Alg. 1 lines 9–11; Φ_LLM). Yet the only ablation (Table 2) is a category breakdown on NExT-QA, not a controlled removal of Think/Reflect/queue updates versus a fixed pipeline or single-pass tool-augmented baseline that uses the same YOLO26, PRTReID, UniSoccer, and SoccerWiki tools. Gains over SoccerAgent (e.g., Q5 100% vs 85.7%, Q14 67% vs 55.5%) are therefore confounded with stronger or better-routed domain tools and RAG. Qualitative failures (Fig. 3b) already show early grounding errors propagating; without a same-tool fixed-plan control, it remains unproven that adaptive DFS systematically reduces rather than merely inherits tool reliability. The NExT-QA 74.16% result further weakens the soccer-specific tree story if tools are largely unused there.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"TreeSoc reformulates soccer video question answering as hierarchical state-space search rather than single-pass VLM prediction. An MLLM coordinator (Qwen3.5-9B) builds a query-specific reasoning tree via dynamic DFS (Alg. 1), decomposing non-atomic queries into ordered subtasks bounded by D_max and W_max, routing leaf nodes to domain tools (YOLO26, PRTReID, UniSoccer, face recognition, camera/foul modules) or retrieval bases (SoccerWiki, SoccerReplay1988, SoccerNet-Caption), accumulating evidence, and applying Reflect/Update-Queue actions (Keep/Modify/Insert/Skip/Retry/Stop) before LLM aggregation. On SoccerBench public test the system reports 85.2% TextQA, 87.4% ImageQA and 82.2% VideoQA (Table 1), outperforming commercial APIs, open VLMs and SoccerAgent on most subtasks; zero-shot NExT-QA accuracy is 74.16% (Fig. 2, Table 2). Code is released.","tokens_in":11841,"tokens_out":1155,"duration_ms":9749,"significance":"If the adaptive tree mechanism is genuinely responsible for the gains, the work supplies a concrete, reproducible agentic template for domain-specific video QA that unifies multi-step decomposition, tool synergy and intermediate replanning. The SoccerBench margins over both general VLMs and the prior soccer agent, together with competitive NExT-QA transfer and public code, would make TreeSoc a useful reference system for sports analytics and tool-augmented video reasoning. Even if gains prove partly tool-driven, the engineering integration and explicit DFS formulation remain of practical interest to the community.","major_comments":[{"comment":"The central claim that dynamic DFS + Reflect/Update-Queue (Alg. 1 lines 9–11; Φ_LLM) is what establishes the “effective paradigm” is not isolated experimentally. Table 1 and §5.2 compare against SoccerAgent and monolithic VLMs, but there is no same-tool control that freezes the plan (fixed queue, no Modify/Insert/Retry) or collapses to single-pass tool-augmented inference while holding YOLO26, PRTReID, UniSoccer, face recognition and the RAG bases fixed. Gains on Q5 (100% vs SoccerAgent 85.7%) and Q14 (67% vs 55.5%) are therefore confounded with tool quality and routing. Table 2 is only a NExT-QA category breakdown, not a mechanism ablation. Without this control the attribution of SOTA performance to adaptive tree search remains unproven; the qualitative failures in Fig. 3b already illustrate early grounding errors that propagate.","section":null},{"comment":"SoccerBench evaluation uses the 500-sample public test (§5.1) with no error bars, bootstrap intervals or significance tests (Table 1). Several baselines leave Q7 blank, and the paper itself notes label noise on Q13. Given the small N and the large claimed margins, statistical reliability of the ranking (especially the 82.2% VideoQA figure) cannot be assessed; at minimum variance estimates or a larger private-test comparison should be supplied before the SOTA claim is treated as settled.","section":null},{"comment":"Cross-domain evidence is thin relative to the paradigm claim. NExT-QA 74.16% (Fig. 2) is competitive, yet §4.2 and the tool bank are heavily soccer-specific; the manuscript does not state which tools (if any) remain active on NExT-QA or whether the DFS planner largely reduces to ordinary multi-step LLM reasoning. If soccer tools are unused, the transfer result supports general agentic decomposition more than “tool-augmented tree reasoning for soccer,” weakening the Abstract/§6 framing.","section":null}],"minor_comments":[{"comment":"Eq. (1) and the surrounding text in §3 introduce c_n = IsSimple ∨ (d_n = D_max) but never define the IsSimple predicate or how the LLM implements it; a short operational description would help reproducibility.","section":null},{"comment":"Figure 1 is useful but the “Self-Correcting Feedback Loop” and “Reasoning Graph” boxes are not mapped to concrete lines of Alg. 1; a brief caption cross-reference would clarify.","section":null},{"comment":"§4.2 notes substantial label noise on Q13 yet still reports the number in Table 1 without a noise-adjusted or human-verified subset; either flag the score more prominently or re-evaluate on cleaned labels.","section":null},{"comment":"Implementation details (§5.1) list decoding hyperparameters and YOLO threshold but omit the concrete values of D_max and W_max used for the reported runs; these free parameters should be stated.","section":null},{"comment":"Minor typography: “T reeSoc” spacing artifacts appear in Table 1 and Table 2 headers; “LLAVA” vs “LLaVA” inconsistency; arXiv date stamp “13 Jul 2026” looks anomalous.","section":null}],"recommendation":"major_revision","confidential_remarks":"The work is solid systems engineering with a clear leaderboard win, but the missing same-tool ablation is load-bearing for any claim that “tree-structured dynamic reasoning” itself is the advance rather than better tool packaging. I would accept after that control (and basic variance numbers) even if the adaptive component proves only modestly helpful; without it the paper risks overselling the search formalism. Scope is appropriate for a CV/multimedia venue that values agentic video systems."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"TreeSoc is a practical multi-step agent for soccer VQA: an LLM coordinator builds a DFS subtask tree, routes to domain tools (YOLO26, PRTReID, UniSoccer, face ID, SoccerWiki/RAG), reflects, and can rewrite the remaining queue. That is not a new principle—tree search and tool agents already exist—but it is a clean, domain-grounded system result with code promised.\n\nWhat it does well is the engineering and the numbers. On SoccerBench public test it posts 85.2 / 87.4 / 82.2 on Text / Image / Video, beating commercial APIs, open VLMs, and SoccerAgent on several hard visual/temporal items (perfect Q5, strong Q14, 100% commentary). NExT-QA at 74.16% zero-shot is a real transfer check, not just soccer tooling. Problem formulation, Alg. 1, and the tool-routing sections are readable; qualitative successes match the claimed pipeline. Citations cover SoccerNet, UniSoccer, VARS/X-VARS, SoccerAgent, and general video agents without obvious gaps. Circularity is low: fixed public MCQs.\n\nSoft spots are real but proportionate. SoccerBench public set is only 500 samples; no error bars or significance tests; some baselines leave Q7 blank while TreeSoc reports it. The stress-test lands: there is no same-tool fixed-plan or single-pass control that isolates Think/Reflect/Update-Queue from better detectors and RAG. Table 2 is a category breakdown, not a mechanism ablation. Failures in Fig. 3b already show early grounding errors propagating, so the claim that adaptive DFS systematically reduces error is still an assumption. Free parameters (D_max, W_max, decoding, YOLO threshold) are ordinary but unanalyzed. Overhead of tree search is acknowledged but not measured.\n\nWho it is for: people building sports analytics agents or tool-augmented video QA who want a concrete recipe and leaderboard evidence. Not for anyone hunting a foundational reasoning theorem. I would send it to peer review; a serious referee can demand the missing same-tool ablation and uncertainty numbers. Worth engaging if the code and tools reproduce.","headline":"Solid soccer VQA agent with real leaderboard gains and shipped code; the DFS/replanning story is under-ablated and gains are partly tool quality.","tokens_in":12467,"tokens_out":539,"would_cite":true,"duration_ms":4898,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"TreeSoc turns soccer video question answering into a dynamic depth-first tree search that routes specialist tools at each node, outperforming single-pass vision-language models.","keywords":["Video Question Answering","Soccer Video Understanding","Tree-Search Reasoning","Multimodal Large Language Models","Multi-Agent Systems","Tool-Augmented Reasoning","Depth-First Search"],"falsifier":"An ablation that freezes the subtask queue after the first decomposition (no Reflect or Update-Queue) and still matches or exceeds full TreeSoc accuracy on SoccerBench VideoQA and NExT-QA would falsify the claim that adaptive replanning is essential.","tokens_in":12405,"feed_emoji":"⚽","tokens_out":767,"duration_ms":16277,"temperature":0.7,"pith_summary":"Contemporary vision-language models still answer complex soccer questions in one forward pass, so early visual mistakes travel straight to the final answer and specialized tools cannot be coordinated on the fly. TreeSoc reframes the same task as hierarchical search: a language-model coordinator recursively breaks a query into an ordered queue of subtasks, walks that tree depth-first, and at every leaf either calls a domain tool (detector, re-identifier, action classifier, knowledge base) or expands further. After each result the coordinator reflects and may rewrite, skip, or retry remaining steps, keeping an explicit evidence state that later nodes can use. On SoccerBench the method records the highest reported scores across text, image and video questions; the identical controller, without soccer tools, still reaches 74 percent on the everyday-video benchmark NExT-QA. The practical claim is that making the reasoning tree and the tool calls visible and revisable yields more reliable multi-step video understanding than scaling a monolithic model.","feed_headline":"Tree search plus tools tops single-pass soccer video QA","feed_subtitle":"Ordered subtasks and on-the-fly specialist modules beat larger vision-language models on SoccerBench and transfer to everyday video.","key_machinery":"Dynamic DFS-based reasoning agent (Algorithm 1): at each node the coordinator Thinks, decides Direct versus decompose into a bounded queue, recursively solves non-leaves, executes leaf tools, accumulates evidence, Reflects, and may Update-Queue via the transition Φ_LLM before synthesizing. The explicit tree plus adaptive replanning carries the claim that visible intermediate states and on-the-fly tool routing outperform monolithic VLMs.","core_discovery":"TreeSoc reformulates soccer video question answering as multi-step state-space search over a hierarchical reasoning tree rather than single-pass prediction. A multimodal language model serves as coordinator: it decomposes each complex query into ordered subtasks, traverses them with dynamic depth-first search, and at every node either solves the leaf with a specialized perception or retrieval tool or expands further. Intermediate results update an evidence state that can trigger replanning—keep, modify, insert, skip, retry or stop—before the root aggregates a final answer. Empirically this yields 85.2 percent TextQA, 87.4 percent ImageQA and 82.2 percent VideoQA on SoccerBench plus 74.16 per","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["TreeSoc recasts soccer video QA as DFS over a tool-augmented reasoning tree","Dynamic depth-first subtasks plus specialist tools beat single-pass VLMs on SoccerBench","Hierarchical search with adaptive tool routing reaches 82.2% VideoQA on SoccerBench","Tree-structured decomposition and on-the-fly tools lift soccer and general VideoQA","Ordered subtask tree with evidence-state replanning tops SoccerBench and NExT-QA"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that the specialized tools and the language model’s reflection steps are accurate enough that rewriting the plan after each result systematically reduces, rather than multiplies, early mistakes.","fun_headline_variants_meta":{"raw":{"variants":["TreeSoc recasts soccer video QA as DFS over a tool-augmented reasoning tree","Dynamic depth-first subtasks plus specialist tools beat single-pass VLMs on SoccerBench","Hierarchical search with adaptive tool routing reaches 82.2% VideoQA on SoccerBench","Tree-structured decomposition and on-the-fly tools lift soccer and general VideoQA","Ordered subtask tree with evidence-state replanning tops SoccerBench and NExT-QA"]},"model":"grok-4.5","effort":"low","cost_usd":0.00483,"raw_usage":{"total_tokens":1441,"prompt_tokens":858,"num_sources_used":0,"completion_tokens":97,"cost_in_usd_ticks":48300000,"prompt_tokens_details":{"text_tokens":858,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":486,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":858,"tokens_out":97,"duration_ms":4033,"temperature":1.0,"reasoning_tokens":486,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T07:48:46.294838+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"An ablation that freezes the subtask queue after the first decomposition (no Reflect or Update-Queue) and still matches or exceeds full TreeSoc accuracy on SoccerBench VideoQA and NExT-QA would falsify the claim that adaptive replanning is essential.","supporting_citations":[],"review_version":1}