Towards Temporal Compositional Reasoning in Long-Form Sports Videos

Conference: ECCV 2026
Paper: ECCV
Code: https://github.com/ustiniansy/SportsTime
Area: Multimodal VLM
Keywords: long-video understanding, temporal compositional reasoning, video QA benchmark, temporal anchoring, GRPO

TL;DR

The paper introduces SportsTime, a long-form sports video benchmark with 14K+ open-ended QA pairs and 50K+ step-wise temporal evidence annotations, together with Chain-of-Time Reasoning (CoTR): training with a temporal-reward GRPO (tr-GRPO) that aligns each reasoning step to a verifiable time anchor, and inference with an anchor-observe-infer loop that retrieves local clips around predicted anchors to verify evidence turn by turn, raising the average accuracy on Qwen3-VL-4B from 25.15% to 29.24% and span IoU from 0.26 to 0.58.

Background & Motivation

Sports videos are multimodal content that is highly dynamic, long-form, and dominated by sparse decisive events — understanding how a soccer goal was scored often requires tracing back a long chain of passes and runs that unfolded well before the final shot. Over the past decade, dedicated vision models have pushed player detection and tracking, action spotting, and tactical analysis far ahead; the rise of MLLMs then made unified understanding plausible, bringing open-ended tasks such as commentary generation and video question answering into a single framework. Yet long-horizon reasoning is precisely where current MLLMs are weakest: a model can describe the few seconds it sees, but struggles to thread sparse decisive evidence scattered across a whole match — pieces separated by minutes — into one reasoning chain. The authors call this capability temporal compositional reasoning: locating multiple temporally dispersed pieces of evidence and composing them into an answer.

This failure is not merely a matter of context length. The paper attributes it to two tightly coupled causes. First, high-quality annotations that explicitly capture evidence spanning multiple time ranges are scarce, so long-horizon reasoning receives almost no supervision — existing sports QA data either uses short clips or provides only the final answer, leaving the model with no signal about which seconds to look at. Second, there is a lack of methods that explicitly encourage a model to identify, localize, and justify the specific temporal evidence behind its answer; answer correctness and evidence validity are two different things, and existing objectives only optimize the former. The combination makes models over-rely on language priors and produce plausible answers that are barely grounded in the video — "the ball went in" can be guessed correctly without watching anything, which is exactly where hallucination comes from.

The paper's direction is therefore to make evidence itself a first-class citizen: first build a benchmark in which every reasoning step carries a checkable timestamp or time span, so that "is the justification right" becomes directly measurable; then design a training objective and an inference procedure that constrain both the answer and the anchor, forcing the model to find evidence in the video rather than patterns in language. Core idea: formulate long-video QA as chain-of-time reasoning, where every intermediate statement must carry a retrievable time anchor — during training, coverage and correctness rewards over anchors (tr-GRPO) make anchors learnable; during inference, predicted anchors serve as retrieval commands for local clips that are verified turn by turn, turning temporal compositional reasoning from an unsupervised latent behavior into an explicit, verifiable, supervised trajectory.

Method

Overall Architecture

The paper contributes both a benchmark and a method. SportsTime supplies the task definition and the supervision signal: each sample is a long video, an open-ended question, an open-ended answer, and a step-wise reasoning chain in which every intermediate step is annotated with the timestamp or time span that supports it (the Chain-of-Time). Three kinds of temporal elements must be composed: the order of events (which step precedes which), the causal chain across events (why a later event follows from an earlier one), and the duration and relative spacing of events (which segment is the decisive instant, how far apart two pieces of evidence are). Because every step carries an anchor, whether a model genuinely composes multiple pieces of evidence becomes measurable step by step: answer accuracy measures the outcome, while the Step-wise Grounding Alignment (SGA) evaluation measures the evidence itself.

CoTR then specifies how a model should operate under that supervision, in three stages: make time visible (visual time anchoring, burning timestamps into the frames), make the model produce anchors (tr-GRPO with temporal rewards), and at inference let the model go look with its anchors (anchor-triggered interactive observation, closing a loop of "generate anchor → retrieve local clip → verify and revise → regenerate"). The first two stages change the model's behavioral tendency; the third changes its input distribution. The three are complementary.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["SportsTime benchmark<br/>14K open QA + 50K time anchors"] --> B["Visual time anchoring<br/>burn mm:ss timestamps into frames"]
    B --> C["tr-GRPO<br/>coverage + correctness rewards"]
    C --> D["Anchor-triggered interactive observation<br/>local clips → verify → revise"]
    D -->|retrieve again with revised anchors| D
    D --> E["answer + verifiable temporal evidence chain"]

Formally, step \(t\) is defined as a pair \(n_t=\langle s_t,\tau_t\rangle\), where \(s_t\) is the textual statement of that step and \(\tau_t\) is its supporting time anchor (a timestamp or a temporal span). The full trajectory is \(\mathcal{T}=[n_1,\dots,n_T,a]\), and the joint probability of answer and chain factorizes into step-wise conditional generation followed by answering conditioned on the whole chain:

\[\pi_\theta(\mathcal{T}\mid V,q)=\Big(\prod_{t=1}^{T}\pi_\theta(n_t\mid V,q,n_{<t})\Big)\cdot\pi_\theta(a\mid V,q,n_{\le T})\]

(⚠️ The cached PDF rendering mangles the tail of this equation; it is reconstructed here from context as answering conditioned on the full chain — refer to the original paper.) The point of this factorization is not mathematical convenience but the interface it imposes: every intermediate statement must carry a retrievable anchor. With that interface, reasoning can go back on-site to verify, and anchors can be rewarded separately during training.

Key Designs

1. The SportsTime benchmark: turning "composition of multiple evidence pieces" into supervised step-wise temporal annotation

The reason long-horizon reasoning is never learned lies in the annotation: telling a model to "integrate multiple pieces of evidence" without saying which seconds they are in leaves the supervision empty. SportsTime writes the evidence anchors explicitly into the annotation and makes three design choices around them. First, open-ended QA instead of multiple choice — closer to real analytical usage, and it closes the shortcut of eliminating candidates, so the output space is the entire answer space. Second, five task types covering different ability levels: perception, temporal, tactical, causal, and counterfactual, ranging from "seeing the frame clearly" up to "what would have happened otherwise"; causal and counterfactual questions are nearly impossible to answer from a single frame and require chaining events that may be far apart. Third, step-wise Chain-of-Time annotation: every intermediate step of the chain carries a timestamp or a second-scale span, which fully decouples "is the answer right" from "was the evidence found", making both directly supervisable and directly evaluable.

In scale and composition, the benchmark covers five team sports — American Football, Ice Hockey, Soccer, Basketball, and Volleyball — which differ substantially in camera conventions, editing styles, and event structures; men's and women's games and international, professional, and collegiate events are included to introduce structured diversity. In total: 1,575 videos (208 full matches + 1,367 highlight videos), average duration 1053.25 seconds, 14,326 open-ended QA pairs, and 50,000+ step-wise temporal evidence annotations. Full matches probe long-horizon reasoning under extremely sparse evidence, while highlights improve coverage of event-dense and stylistically diverse scenarios.

The annotation pipeline is an expert-guided semi-automatic three-stage process, motivated by a practical constraint: long-form sports QA simultaneously demands scale, temporal validity, and domain correctness — purely manual construction is unaffordable and purely automatic is untrustworthy. In stage one, sports domain experts design question templates that fix the question types and semantic scope, converting open-ended generation into controlled generation. In stage two, a strong MLLM instantiates the templates against the videos to produce candidate QA samples and their temporal evidence, which supplies the scale. Stage three is a two-stage manual review: peer annotators perform an initial check on format consistency, temporal validity, and clarity of expression to filter obviously low-quality samples, after which domain experts do the final review, focusing on rule correctness, tactical plausibility, and the feasibility of counterfactual scenarios. The pipeline keeps the efficiency of scaled generation while suppressing temporal errors, semantic ambiguity, and domain-knowledge bias.

2. Visual time anchoring: let the model read the time instead of inferring it

A very practical obstacle is that even when a model correctly identifies an event, aligning it to an exact time still drifts. The cause is not visual recognition but that time is not a directly observable quantity in video: frame indices, sampling rates, and context length all make "which minute is this" something to be inferred, inference error accumulates over long contexts, and anchors shift systematically. Inspired by the "a picture is worth a thousand words" philosophy behind DeepSeek-OCR and Number_it, the authors use a nearly cost-free trick: burn an explicit timestamp in a consistent mm:ss format into the top-right corner of every frame, so the video carries its own clock.

This downgrades temporal grounding from a reasoning problem to visual-text recognition: the model no longer has to reconstruct a timeline from a long context, it just reads the digits in the frame, and anchor drift over long contexts drops accordingly. The approach is plug-and-play — no extra parameters and no extra training objective — and orthogonal to the later reward design; the costs are mild frame corruption and a dependence on a fixed time format. In the ablation, removing the timestamp overlay drops overall accuracy from 27.31% to 26.12%, confirming that this "simplest possible trick" genuinely matters. It is also a precondition for the third stage: only when the model reliably emits a parseable mm:ss can an anchor be used as a retrieval command.

3. tr-GRPO: replacing the sparse answer reward with coverage and correctness rewards over time

With only answer correctness as reward, a model can collect reward by "giving no time anchor at all and still guessing the answer". This is exactly why generic RL fine-tuning brings only a marginal gain and even hurts temporal grounding (in the ablation, Native-GRPO drops Temporal from 13.14 to 11.76): it puts no pressure on where the evidence comes from. CoTR adds a term that specifically rewards temporally grounded reasoning. The total reward is a weighted sum of three components:

\[R(\mathcal{T})=\lambda_{\text{fmt}}r_{\text{fmt}}(\mathcal{T})+\lambda_{\text{acc}}r_{\text{acc}}(\mathcal{T})+\lambda_{\text{temporal}}r_{\text{temporal}}(\mathcal{T})\]

where \(r_{\text{fmt}}\) is a binary structural reward (whether the output follows the required format) and \(r_{\text{acc}}\) is a task-level answer correctness reward. The focus is \(r_{\text{temporal}}\), itself split into two complementary halves:

\[r_{\text{temporal}}(\mathcal{T})=\alpha\,r_{\text{cov}}(\mathcal{T})+(1-\alpha)\,r_{\text{cor}}(\mathcal{T})\]

The coverage term \(r_{\text{cov}}\) segments the model's <thinking> into steps and computes the proportion of steps that contain at least one explicit time anchor. It guards against omission: the model skipping a piece of evidence and continuing to reason with no anchor for it. The correctness term \(r_{\text{cor}}\) extracts the anchors predicted inside <thinking> and matches them against the ground-truth step anchors: for each ground-truth anchor it computes a best-match score against the predicted anchor set, using span IoU for span-span matches and a distance-aware similarity for point-span or point-point matches, then averages over all ground-truth anchors to land in \([0,1]\). It guards against drift: emitting many timestamps that are attached to the wrong moments. (The values of the \(\lambda\) weights and of \(\alpha\), along with the detailed matching rules, live in Appendices B.2 / A.3 and are not given in the main text — ⚠️ refer to the original paper.)

Together the two terms turn a time anchor from a decorative output fragment into a rewarded intermediate state: \(r_{\text{cov}}\) ensures evidence is covered, \(r_{\text{cor}}\) pulls predicted anchors onto the annotated evidence, and the policy therefore prefers reasoning chains whose anchors align with verifiable evidence. GRPO is used rather than plain policy gradient because group-relative advantages are more stable for long chains with sparse rewards: several trajectories are sampled for the same \((V,q)\), rewards are computed, and rollouts are compared within the group, which avoids the variance of absolute reward scale. The effect is clearest in SGA — tr-GRPO reaches 95.76% anchor coverage with mIoU 0.5812 and [email protected] 56.89%, whereas Native-GRPO reaches only 13.10% coverage and 0.3128 mIoU: their answer accuracies are close (27.31 vs 26.72), but the evidence quality differs by an order of magnitude.

4. Anchor-triggered interactive observation: turning predicted anchors into "go back and look" retrieval commands

Training makes the model willing to emit anchors, but a chain generated in a single forward pass can still be wrong, because under a limited frame budget the model may simply never have sampled the decisive instant. CoTR closes that gap at inference: since the model has already produced a trajectory \(\mathcal{T}=[n_1,\dots,n_T,a]\) with anchors, each predicted anchor \(\tau_t\) is treated as an explicit temporal query that retrieves local evidence from the corresponding neighborhood of the video. For a point anchor \(\tau_t=\texttt{mm:ss}\), a short window centered on it is sampled; for a span anchor \(\tau_t=[t_s,t_e]\), several clips are sampled uniformly within the span. Each anchor is thus converted into a set of local clips, each clip consisting of \(L\) frames sampled at a fixed stride. This replaces global scanning of a long video with bounded local observation around the predicted evidence locations, keeping the cost controllable.

Given the clips, the model performs evidence-grounded reasoning: at turn \(t\) it receives the question \(q\), the current step statement \(s_t\), the anchor \(\tau_t\), and the retrieved local clips, and uses this real footage to verify and revise the current step before moving to the next turn. Once all turns are done, the final answer is generated conditioned on the accumulated, verified evidence. This addresses something reward-based training alone cannot: rewards shape behavior on the training distribution, whereas the loop makes the model re-check each intermediate claim against the actual frames of this particular video. Accordingly, adding AT-IO lifts the overall score further from 27.31% to 29.24% (Temporal 13.73 → 15.43), showing that test-time verification and training-time reward are complementary rather than redundant.

A Worked Example

(Note: the paper does not provide a turn-by-turn example; the walkthrough below is constructed from the mechanisms above to show how state changes across turns.) Take the question "Why did the defense fail on this play?" In the first turn the model proposes a speculative step \(s_1\) = "the defense suffered a matchup error in midfield" with anchor \(\tau_1\) = 12:34. AT-IO samples a window centered at 12:34, the model sees that the matchups were in fact intact, and therefore revises the step to "the defense still held its shape at 12:34" while moving the anchor to 13:02, the moment the mismatch actually occurs. In the second turn the model retrieves footage around 13:02 to confirm the mismatch, then traces backward to a substitution at 11:47, chaining "substitution → matchup mismatch → defense broken" and finally answering conditioned on these three verified pieces of evidence. The key throughout is that the anchor is not decoration for the explanation but the input to the next retrieval: a wrong anchor retrieves the wrong footage and gives the model a chance to self-correct. That is precisely the mechanism by which "composing multiple pieces of evidence" becomes explicit rather than latent.

Loss & Training

The backbone is Qwen3-VL-4B, with 128 frames sampled per video for training and up to 768 frames at inference; all experiments run on 2× NVIDIA H100 (80GB). The objective is the tr-GRPO goal described in design 3: for each \((V,q)\), several anchored trajectories are sampled, rewards \(R(\mathcal{T})\) are computed, group-relative advantages are formed, and the policy is updated starting from \(\pi_0\) while preserving the anchored generation protocol. One notable negative result: direct SFT is worse than the base model (25.15% → 17.61%), which the authors attribute to the structural variability and non-uniqueness of temporal reasoning chains — the same video admits several valid evidence compositions and orderings, token-level imitation of such multimodal outputs is brittle at the 4B scale, and generation becomes highly repetitive. This is the direct justification for reward shaping over imitation.

Key Experimental Results

Main Results

Main results on SportsTime (open-ended QA, LLM-as-Judge with a fixed Qwen2.5-VL-7B judge; all scores in %). The ceiling is strikingly low:

Model	Visual Input	Perception	Temporal	Tactical	Causal	Counterfactual	Avg.
GPT-5	0.2 fps	38.77	27.45	43.85	44.77	46.93	40.72
Gemini-2.5-Pro	0.2 fps	39.66	19.78	29.49	39.41	13.71	29.37
Qwen3-VL-8B-Instruct	768 frames	24.27	14.31	33.26	23.08	44.47	27.45
Qwen3-VL-4B-Instruct (base)	768 frames	23.26	13.14	31.11	22.92	34.17	25.15
Video-R1-7B	768 frames	19.26	8.04	23.25	18.62	33.01	20.40
VideoLLaMA3-7B	768 frames	20.06	10.98	16.92	14.00	28.54	17.96
InternVideo2.5-8B	512 frames	18.12	10.39	19.83	9.85	26.21	16.68
MiniCPM-V4.5-8B	512 frames	16.38	10.14	13.50	16.18	27.27	16.48
GLM-4.6v-Flash-9B	640 frames	2.27	1.18	3.76	3.54	13.20	4.62
Ours (tr-GRPO)		22.01	13.73	35.73	25.54	39.81	27.31
Ours (full)		26.42	15.43	34.55	27.21	42.22	29.24

CoTR's absolute gains over the base Qwen3-VL-4B are: Perception +3.16, Temporal +2.29, Tactical +3.44, Causal +4.29, Counterfactual +8.05, overall +4.09; on a 200-example stratified human-evaluated subset the overall score rises from 24.50% to 30.50% (+6.00), indicating that the LLM judge's ranking agrees with human judgment. Transfer to general long-video benchmarks is consistently positive relative to the same base model: LVBench 56.2 → 59.9 (+3.7), VideoMME 69.3 → 72.1 (+2.8), MLVU 75.3 → 76.2 (+0.9).

Ablation Study

Component ablation on SportsTime (all scores in %):

Config	Temporal	Tactical	All
Base (Qwen3-VL-4B-Instruct)	13.14	31.11	25.15
+ SFT	6.98	19.90	17.61
+ Native-GRPO	11.76	32.65	26.72
+ tr-GRPO w/o ts (no timestamp overlay)	13.14	32.21	26.12
+ tr-GRPO w/ ts	13.73	35.73	27.31
+ AT-IO (Ours, full)	15.43	34.55	29.24

Key Findings

• SGA shows that "asking for timestamps" and "using timestamps" are two different things — the most compelling contrast in the paper:

Model	Acc(%)↑	Anchor(%)↑	mIoU↑	[email protected](%)↑
Base	25.15	11.95	0.2601	30.03
Zero-shot CoT	24.01	19.49	0.2042	26.34
Time-prompted CoT	19.53	91.80	0.1244	12.12
Native-GRPO	26.72	13.10	0.3128	32.09
tr-GRPO (Ours)	27.31	95.76	0.5812	56.89

Merely prompting for timestamps pushes anchor coverage to 91.80%, yet mIoU falls to 0.1244 and [email protected] is only 12.12% — the model learns format compliance and attaches arbitrary time windows to intermediate claims to look grounded. tr-GRPO reaches an even higher coverage of 95.76% while lifting mIoU to 0.5812, showing that reward shaping changes evidence faithfulness rather than output format. - SFT is a negative result: direct supervised fine-tuning drops to 17.61%, below the untrained base, with heavy repetitive generation. The cause is that temporal reasoning chains are structurally variable and non-unique, so token-level imitation is brittle and hard to optimize on a multi-solution task — a negative result that itself argues for reward shaping over imitation on evidence-chain tasks. - The components are complementary rather than redundant: tr-GRPO contributes +2.16 (25.15 → 27.31) and AT-IO a further +1.93 (27.31 → 29.24), with AT-IO mainly helping Temporal (13.73 → 15.43). Removing the in-frame timestamp overlay costs a net 1.19 points, confirming that making time a readable signal is a precondition for the temporal reward to work. - The largest gains are on Counterfactual (+8.05) and Causal (+4.29), exactly the two task types that depend most on composing evidence across events; Perception also improves (+3.16), suggesting the method improves both finding the relevant visual evidence and integrating dispersed clues into causal attribution. - Longer video does not mean harder: accuracy does not decrease monotonically with video length. The authors argue difficulty is driven by confounding factors such as video type and event structure rather than length itself, and CoTR retains its advantage on longer videos. - Judge reliability: three LLM judges (GLM-4.7, MiniMax-M2.5, Qwen2.5-VL-7B) achieve 88.34% average pairwise agreement with Fleiss' κ = 0.57 (moderate consistency); Cohen's κ against human judgment is 0.6467 / 0.5759 / 0.5882 respectively. This supports LLM-as-Judge for scalable evaluation, though agreement is not strong. - Human SGA assessment shows the chains score best on faithful and logical consistency (FLC 4.50/5) and are largely reliable at the intermediate-claim level (ICR 4.02/5), but weaker on evidence-span accuracy (ESA 3.65/5) — the chain reads coherently, yet the exact seconds are not pinned down, consistent with the objective mIoU of 0.58.

Highlights & Insights

• Burning timestamps into frames is the best value-for-effort trick in the paper: zero training cost, zero parameters, simply converting time from a latent variable to be inferred into an explicit signal to be recognized. Any task that requires a model to output a time or a location (temporal localization, citation, region selection) can borrow the idea — make the thing being referenced visible in the input.
• The "ask for timestamps ≠ use timestamps" contrast is a beautifully designed experiment: time-prompted CoT pairs high coverage with low mIoU, showing crisply that prompting buys format but not faithfulness. It is clean empirical evidence that evidence quality must be measured separately, and it explains why two reward terms (coverage and correctness) are needed rather than one.
• The coverage/correctness reward decomposition transfers to any evidence-chain supervision: coverage prevents omission, best-match similarity prevents misplacement — a binary pattern that applies equally to citation quality in retrieval-augmented generation, tool-call traces in multi-step agents, and rationale supervision in interpretability work.
• Anchors as retrieval commands (AT-IO) turns a generated intermediate result back into the input of the next turn, forming a generate-retrieve-verify loop. It needs no external retriever; the retrieval range is entirely determined by anchors the model itself proposes, making it a very lightweight form of test-time verification.
• Benchmark and method are mutually enabling: without step-wise temporal annotation, \(r_{\text{cor}}\) cannot be defined; without \(r_{\text{cor}}\), the failure mode of "attaching arbitrary time windows" would never be exposed. This kind of "data design makes a certain supervision possible" combination is the most solid form a dual-contribution paper can take.

Limitations & Future Work

• Limitations acknowledged by the authors: fine-grained temporal localization remains improvable (ESA only 3.65/5, mIoU 0.5812), with predicted spans often directionally correct but boundary-imprecise.
• ⚠️ Limitations I notice: the main text does not report the values of \(\lambda_{\text{fmt}}/\lambda_{\text{acc}}/\lambda_{\text{temporal}}\) or \(\alpha\), nor any sensitivity analysis (they are deferred to the appendix), so a reader cannot tell whether the temporal reward weighting is sensitive — yet this is the method's most critical knob.
• The method is validated only on Qwen3-VL-4B, with no larger-backbone or proprietary-model augmentation experiment. GPT-5 still reaches 40.72% on SportsTime while the CoTR-enhanced 4B model reaches 29.24%; whether the method meaningfully narrows that gap, or stacks onto 8B/30B models, is not answered.
• "Multiple pieces of evidence must be composed; a single frame cannot suffice" is currently supported by data design intent, the two-stage human review, and the SGA metric, but there is no quantitative shortcut control: no single-frame oracle upper bound, no performance under shuffled timestamps, and no perturbation study measuring how much removing the decisive evidence clip costs. Without these, the necessity of composition remains a claim rather than a measurement.
• Human evaluation is small: 200 examples for open-ended QA and 100 for SGA (five raters). The inference cost of AT-IO (total frames, latency, and memory from multi-turn local retrieval) is not reported, even though inference budget is a key cost item for long-video methods.
• The data covers five sports with a North American professional/collegiate and soccer emphasis; transfer to other sports (individual events, combat sports, athletics) and other broadcast cultures is unknown, and the English commentary and rules context limits cross-lingual applicability.
• Improvement directions: replace \(r_{\text{cor}}\)'s per-ground-truth best-match with a sequence alignment (e.g. DTW) to reward order correctness — as designed, a chain whose anchor set is correct but whose order is reversed receives no penalty, even though order is central to causal reasoning; further, jointly reward time anchors with player/spatial anchors to move from "which second" toward "which second, who, and where".

Related Work & Insights

• vs sports QA benchmarks such as SportQA / Sports-QA / SPORTU / SPORTR / DeepSport: these focus mostly on short clips or fine-grained action labels (durations from seconds to clip level), and although SPORTU / SPORTR / DeepSport introduce CoT supervision, none aligns reasoning steps to a timeline. SportsTime differs by being long-form (average 1053 seconds), fully open-ended, and — crucially — by annotating every reasoning step with a timestamp or second-scale span, making intermediate paths directly supervisable and evaluable.
• vs VRBench: VRBench also provides CoT and step-wise time annotation, but targets narrative videos where spans are often minute-scale. Decisive sporting events are momentary, so SportsTime demands second-level timestamps or short spans; the localization precision requirement is far stricter, which is why this method must handle anchor drift at all.
• vs TOGA: TOGA jointly predicts an open-ended answer and the temporal spans supporting that answer — essentially a single-step map from question to answer plus evidence set. CoTR upgrades this into a multi-step chain where anchors become intermediate states between steps, used both for training supervision and for test-time evidence acquisition.
• vs LongVT / VITAL: both use temporal localization as part of a tool-augmented reasoning process in which the model calls tools to inspect the video. CoTR's anchors serve a similar retrieval-command role, but are generated naturally within the reasoning chain with no separate retriever or tool interface, making the approach lighter; the cost is that retrieval quality depends entirely on anchor quality.
• vs general video RL fine-tuning such as Video-R1 / Native-GRPO: under generic rewards these improve answer accuracy while barely improving temporal grounding (coverage 13.10%, mIoU 0.3128). The lesson is that when the bottleneck is where the evidence comes from rather than whether the answer is right, the evidence itself must be written into the reward, otherwise RL only reinforces the language priors already present.

Rating

• Novelty: ⭐⭐⭐⭐ The benchmark-plus-method combination is complete (step-wise temporal annotation + temporal reward + anchor loop), though the individual techniques have precedents (in-frame timestamp overlays appear in Number_it; GRPO is a general framework); the novelty lies mainly in formalizing time anchors as supervised intermediate states.
• Experimental Thoroughness: ⭐⭐⭐⭐ Main results, dual-track SGA (objective and subjective), component ablation, video-setting ablation, transfer to general benchmarks, and a multi-judge reliability study are all present; however only a 4B backbone is validated, there is no quantitative single-frame shortcut control, AT-IO inference cost is unreported, and key reward hyper-parameters sit in the appendix.
• Writing Quality: ⭐⭐⭐⭐ The argument chain is clear and the "high coverage, low alignment" contrast of time-prompted CoT is the highlight; the drawbacks are mangled equations in the rendering and implementation details (reward weights, anchor parser) deferred to the appendix, weakening self-containment.
• Value: ⭐⭐⭐⭐ Provides both a supervisable benchmark and transferable reward/inference paradigms for evidence-chain understanding in long video; the coverage-plus-correctness reward decomposition and "anchors as retrieval commands" are directly applicable to long-video QA, citation-quality supervision, and multi-step agent traces.

Related Papers

• [ACL 2026] TemporalVLM: Video LLMs for Temporal Reasoning in Long Videos
• [ECCV 2026] VideoTIR: Accurate Understanding for Long Videos with Efficient Tool-Integrated Reasoning
• [ICLR 2026] TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning
• [ECCV 2026] Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding
• [ECCV 2026] Falcon: Functional Assembly and Language for Compositional Reasoning in X-ray