acceptodds
Under review as a conference paper at ICLR 2027

From Scan to Focus: Learning a Value-Guided Adaptive Browsing Policy for Efficient Long-Video Reasoning

Abstract

Long-video reasoning requires Multimodal Large Language Models (MLLMs) to locate sparse evidence over extended temporal contexts while preserving the fine-grained visual details needed for reliable answers. Uniform sparse sampling can miss brief but decisive events, whereas dense processing incurs substantial visual-token and memory costs. To address this challenge, we introduce MindSight, a human-inspired adaptive browsing agent that treats long-video understanding as active perception. MindSight follows a scan–focus–reflect process: it first builds a low-cost global view to identify candidate intervals; conditioned on a candidate interval and the evidence acquired so far, it then compares STOP, LOW, and HIGH by their expected downstream returns, deciding whether to stop or inspect the interval at low or high visual fidelity by trading off new evidence against its observation cost. We learn this observation-value decision through paired Monte Carlo continuations at matched timestamps, and optimize multi-turn browsing trajectories with turn-balanced geometric-mean policy optimization (TB-GMPO). Experiments with two backbones on four long-video benchmarks improve accuracy by 1.2–12.0 percentage points (pp) over the corresponding base models, while reducing visual-token consumption and improving inference efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Multimodal Models;Long video reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_91FZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MindSight first scans a video at low resolution, then selects a temporal interval and chooses STOP, LOW, or HIGH. Paired continuations at matched timestamps provide cost-aware action targets. Replay-verified SFT is followed by TB-GMPO, which balances contributions across turns and bounds log-ratio updates. Experiments on two backbones and four benchmarks report accuracy gains and reduced visual-token use.

Critical strengths
  1. Well-motivated action supervision. Matching timestamps for LOW and HIGH distinguishes visual fidelity from temporal localization. Figure 3 reports comparable or better accuracy than always-HIGH with substantially fewer acquired visual tokens.
  2. Informative ablations and cost accounting. Three-seed component ablations and separate accounting for acquired tokens, replayed prefill, and serving latency make the evaluation informative. The cached and full-replay comparisons are particularly useful for understanding deployment costs.
  3. Coherent training pipeline. The method combines replay-verified trajectories, action targets that account for downstream observation costs while excluding ambiguous comparisons, and turn-balanced policy optimization (Sec. 3.2). These choices address trajectory validity, supervision quality, and uneven turn lengths.
Critical weaknesses
  1. Cost-aware action preferences may shift during RL. Labels are generated with a frozen SFT snapshot, whereas TB-GMPO changes the continuation policy and uses a different reward (Eqs. 3 and 5). The paper acknowledges that targets can become stale, but does not quantify how their cost-adjusted action preferences change. Such changes need not be harmful, but their effect on the final accuracy-cost trade-off is unclear.
  2. The speedup is conditional on cache reuse. Table 3 reports 3.46 seconds to the first final-answer token with prefix caching and 7.92 seconds with full replay, compared with 4.72 seconds for Uniform Sparse. The paper clearly demonstrates this dependence. The remaining question is which cache settings and reuse rates are needed to obtain the speedup in practice.
Questions for authors
  1. On fixed development states, how often does the best action under the cost-adjusted value criterion change from the SFT snapshot to the final RL policy? How do these changes affect the accuracy-cost trade-off?
  2. What cache settings and prefix-reuse rates produce Table 3's results, and under what deployment conditions would the speed advantage diminish? Please clarify applicable caching optimizations for each method, without assuming equal cross-turn reuse for single-pass baselines.
  3. What fraction of paired-supervision states pass filtering, and what is their STOP/LOW/HIGH label distribution? How do actual action frequencies evolve during RL relative to the benchmark proportions already reported in Figure 3?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 91FZ (1/3)

We sincerely thank the reviewer for the careful and constructive assessment. The comments helped us clarify three important issues: cost-adjusted action preference changes after RL, cache requirements and deployment conditions, and the distinction between training-label statistics and inference-time action frequencies. In the revision, we have added the confirmed supervision and budget results and clarified the limitations of the current evidence. For clarity, we respond point by point below.

1. Cost-adjusted action preferences after RL

The reviewer is correct that paired labels are defined relative to the frozen SFT continuation policy that generated them. TB-GMPO then changes this continuation policy by optimizing answer correctness, temporal grounding, and format rewards. Its return does not directly include the acquired-token penalty used in the paired value criterion. Therefore, reference-policy KL alone does not guarantee unchanged action rankings.

Under an identical TB-GMPO training setup, adding paired labels improves accuracy from 38.4/54.2% to 39.5/56.0%, a gain of 1.1/1.8 points on LVBench/VideoMME-Long, respectively. This directly shows that paired supervision remains beneficial in the tested RL pipeline. To answer the reviewer’s question about the frequency and consequences of individual stale labels directly, we re-estimate action values on the same fixed development states, using the same cost-adjusted criterion and confidence rule under both continuation policies. Visible evidence, candidate actions, timestamps, resolutions, and remaining budgets are held fixed.

This uses the same cost-adjusted criterion under each policy, rather than comparing argmax actions under different reward definitions. From the label-generating SFT snapshot to RL initialization, the preferred action changes in 8.9% of 1,012 jointly confident states. From RL initialization to the final RL policy, it changes in 11.7% of 1,084 jointly confident states (Table R5a).

Table R5a. Fixed-state preference audit

Policy comparisonFixed statesJointly confident statesOptimal-action flip (%)Confident→ambiguous (%)Ambiguous→confident (%)Incomplete evaluations
Label-generating SFT → RL initialization1,5001,0128.96.118.743
RL initialization → final RL1,5001,08411.74.816.937

These results exhibit the pattern expected from an RL stage that is both effective and policy-improving. The cumulative flip rate from the label-generating SFT snapshot to the final RL policy is moderate rather than catastrophic. Moreover, ambiguous→confident transitions occur more frequently than confident→ambiguous transitions , suggesting that the final policy does not simply drift away from the original labels. This suggests that the final policy does not simply drift away from the original labels; instead, it makes a substantial subset of previously uncertain states more decisively evaluable.

Table R5b. Consequences under the final continuation policy

SubsetStatesMean old-label value regretCorrectness difference (pp)Acquired visual-token difference
Evaluable states with a clear old label1,3280.00634+0.67+111
Jointly confident, changed action1520.0314+4.6+624
Jointly confident, unchanged action8200.0018+0.1+21
Old label clear, final preference ambiguous3560.0061+0.3+98

These results support two points. First, stale labels can matter: the changed-action subset shows a larger mean old-label value regret ( 0.0314 ) and a larger correctness difference ( +4.6 pp ). Second, the aggregate effect remains limited because changed confident actions account for only a subset of all evaluable states. This is consistent with our main claim: paired supervision remains useful in the tested RL pipeline, while individual stale labels should be audited rather than assumed harmless.

In the updated evaluation reported in our response to Reviewer xjPX, MindSight achieves 48.5% accuracy, outperforming the local official-style Qwen baseline by 3.8 percentage points (44.7%). Compared with fixed-resolution Qwen, MindSight achieves higher accuracy (48.5% vs. 43.8%) with a 23.4% lower maximum visual-token budget (50,176 vs. 65,536), half the context limit (64K vs. 128K), and fewer permitted frame observations (488 vs. 512). These results demonstrate a favorable accuracy–cost trade-off: MindSight delivers better accuracy under tighter visual-token and context budgets.

AuthorsOpenReview ↗

Response to Reviewer 91FZ (2/3)

2. Cache settings, reuse rates, and deployment conditions

Table 3 in the paper was measured on four A100-80GB GPUs, TP=4, batch size=1. MindSight’s prefix-cache setting reuses KV states for the identical multimodal history across turns, whereas full replay recomputes that history.

As Table R6 shown, under this setup, TTFT is 3.46 s with prefix reuse, 7.92 s with full replay, and 4.72 s for Uniform Sparse. Thus, cached MindSight has 26.7% lower TTFT than Uniform Sparse, whereas full replay has 67.8% higher TTFT. MindSight’s latency advantage over Uniform Sparse is conditional on prefix reuse; it reverses in the reported full-replay configuration.

TTFT includes scan, tool, and reasoning calls before the first final-answer token. The measured complete-answer latency is 4.91 s for cached MindSight versus 6.25 s for Uniform Sparse, a 21.4% reduction. These timings belong to the original serving configuration and are not measurements of the updated 488-frame evaluation.

Table R6. Reported serving results and reuse statistics.

MethodTTFT (s)TPOT (ms/token)Complete answer (s)Peak total GPU memory (GB/GPU)Intra-episode KV hit rate (%)Mean executed prefill tokens per episode
MindSight, prefix cache3.4628.44.9132.164.818,432
MindSight, full replay7.9230.89.8441.20 (reuse disabled)39,936
Uniform Sparse4.7229.16.2528.5N/A24,576
Uniform Dense16.9029.718.4264.2N/A61,440

Intra-episode KV hit rate is the ratio of eligible history tokens actually reused in later turns to all eligible history tokens in those turns, excluding newly appended observations and text. Executed prefill tokens are summed within each episode and then averaged across episodes. Peak total GPU memory denotes the peak total memory occupied on each GPU, including allocated KV memory, model weights, and other runtime allocations.

The reuse statistics show why the latency gain is plausible under the reported serving setup. The cached configuration reuses a substantial fraction of eligible history states and executes much less prefill work than full replay. At the same time, the comparison does not assign MindSight’s intra-episode reuse opportunity to the single-pass baselines. For Uniform Sparse and Uniform Dense, intra-episode KV hit rate is marked as N/A, because those methods do not have the same multi-turn browsing structure.

Effective reuse requires an unchanged serialized multimodal prefix and KV retention or availability across turns. The advantage can diminish under cache-capacity pressure, concurrent eviction, prefix changes, or routing to a worker without the required KV states. Therefore, the revised manuscript limits the speed claim to the measured serving conditions and defines reuse statistics separately from visual-encoding reuse and cross-question reuse.

Table R7. Method-specific optimization scope and configuration record

MethodIntra-episode reuseAdditional applicable optimizationEngine/versionInitial cache stateRouting/eviction
MindSightIdentical history KV across browsing turnsCommon text-prefix and repeated visual-encoding cache where supportedvLLM 0.6.0Cold at request startSticky routing, LRU eviction
Uniform SparseN/A: single passCommon text-prefix and repeated-video/cross-question reuse where supportedvLLM 0.6.0Cold at request startSticky routing, LRU eviction
Uniform DenseN/A: single passCommon text-prefix and repeated-video/cross-question reuse where supportedvLLM 0.6.0Cold at request startSticky routing, LRU eviction

Under the reported serving configuration, MindSight achieves lower latency than Uniform Sparse with cross-turn prefix reuse, but higher latency without it. The speed claim is therefore conditional on effective KV retention and reuse.

AuthorsOpenReview ↗

Response to Reviewer 91FZ (3/3)

3. Training-label acceptance and actual action frequencies during RL

Of 6,840 attempted states , 6,651 completed all action evaluations and 4,218 received clear labels. Among the completed states, 2,433 were ambiguous and 189 were incomplete. The acceptance rate is therefore 61.7% of attempted states , or 63.4% of completed states.

Table R8a. Accepted training labels

ActionLabelsRate (%)
STOP1,77242.0
LOW1,40133.2
HIGH1,04524.8

The labeling configuration uses 32 continuations per action , λC=0.10 , τV=0.02 , and 1,000 bootstrap resamples , with nominal 95% simultaneous confidence after Bonferroni correction. Ambiguous states contribute no action-label loss. Figure 3 in the paper describes final-policy inference actions under the original evaluation budget. It uses different populations and denominators from Table R8a, so the two tables should not be directly interpreted as a drift estimate.

Table R9. Checkpoint action audit on the common held-out RL validation set

CheckpointValidation accuracy (%)Parseable decisionsSTOP/LOW/HIGH (%)LOW/HIGH calls per episodeAcquired visual tokensExecuted prefill tokensFalse STOP (%)
Label-generating SFT snapshot55.81,18741.8 / 35.4 / 22.81.8417,92031,48811.2
Post-label SFT / RL initialization60.01,19139.9 / 36.0 / 24.11.9718,78430,1129.4
Intermediate RL66.11,19337.6 / 36.8 / 25.62.1119,96828,7367.1
Final RL, 10k updates72.51,19436.9 / 37.1 / 26.02.1920,35227,9045.8

Table R9 tracks actual action frequencies across checkpoints on the same held-out validation questions under a fixed evaluation budget . During RL, the action shares change consistently from initialization through the intermediate checkpoint to the final checkpoint:

  • STOP decreases: 39.9% → 37.6% → 36.9% ( −3.0 percentage points overall).
  • LOW increases: 36.0% → 36.8% → 37.1% ( +1.1 percentage points ).
  • HIGH increases: 24.1% → 25.6% → 26.0% ( +1.9 percentage points ).

Together, LOW/HIGH account for 60.1% → 62.4% → 63.1% of parseable decisions , while their combined calls per episode rise from 1.97 to 2.19 (+0.22) . Thus, RL shifts the policy moderately toward additional observations rather than earlier stopping. This shift accompanies an increase in validation accuracy from 60.0% to 72.5% (+12.5 percentage points) and a decrease in False STOP from 9.4% to 5.8% (−3.6 percentage points).

We hope our responses have addressed your concerns and clarified the relevant aspects of our work. Please feel free to raise any additional questions; we would be glad to provide further clarification.

We would be grateful if you would consider reassessing our work in light of these clarifications.

Thank you for the clarifications. The new analyses address my main concerns, I would like to raise my score from 2 to 3.

Reviewer_xjPXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This article proposes MindSight, a multi round browsing agent for long video Q&A. The process is divided into three steps: first, use low resolution global scanning to propose candidate intervals; Then select from the three actions of STOP/LOW/HIGH for each candidate interval; Finally, decide whether to continue based on new evidence. Training is divided into three stages: (1) Bootstrap and replay on LongVideo Reason to verify about 3K trajectories for SFT; (2) Using the frozen SFT strategy, perform three actions in the same state, each with M=32 Monte Carlo sequels, and give a hard label by subtracting the visual token cost from the accuracy; (3) Propose TB-GMPO (GMPO+round wise weighting+tanh logarithmic ratio constraint) for multi round RL. Report a relative base improvement of 1.2% -12% on two backbones and four long video benchmarks, and provide token and A100 inference efficiency analysis

Critical strengths
  1. The research question is clear and has a reasonable motivation. MindSight jointly considers time localization, observation accuracy, and stopping decisions, rather than simply treating video browsing as a fixed sampling process. Using the same timestamp for LOW and HIGH can help distinguish the impact of visual observation accuracy.
  2. The discussion on efficiency is quite comprehensive.
Critical weaknesses
  1. The independent contribution and reliability of paired Monte Carlo value supervision have not been fully demonstrated. How many states have obtained reliable action labels, how many states are filtered due to unclear value differences, and whether action labels remain stable as the sampling frequency and value interval change are not seen in the article. The value of paired value supervision is also not significant in ablation experiments.
  2. The baseline of the pedestal model is significantly low, and the credibility of the main conclusion is questionable. In Table 1, Qwen2.5-VL-7B is 38.3% on LVBench, while the official report for Qwen2.5-VL is 45.3%; The results of MindSight-7B are even lower than the official pedestal numbers. Moreover, the paper did not specify how many frames, what resolution, or what context were used in the pedestal evaluation.
Questions for authors
  1. Can the author evaluate the independent contributions of paired value supervision and TB-GMPO under the same training data, reward design, trajectory sampling budget, and random seeds?
  2. How many states were evaluated by paired Monte Carlo simulations? How many of these states have obtained clear action labels? What is the distribution of the STOP, LOW, and HIGH labels?
  3. How many frames, resolutions, and contexts did the pedestal model use under your 'unified protocol'? Why is Qwen2.5-VL-7B almost 7 percentage points lower than the official report on LVBench? Can you provide the results of the pedestal under multiple frame budgets.
2 comments
AuthorsOpenReview ↗

Response to Reviewer xjPX (1/2)

We sincerely thank the reviewer for recognizing MindSight’s motivation, browsing design, and efficiency analysis. We address the three concerns below. Existing ablation results are reported as mean accuracy ± standard deviation over training seeds 42, 100, and 250.

1. Independent contributions of paired value supervision and TB-GMPO

To assess whether each component also benefits the model without the other, we extend the ablations in Table 2 of the paper to a full 2×2 comparison in Table R1 to evaluate each component with and without the other.

Table R1. Paired value supervision × optimizer. Accuracy (%), mean ± SD (pp) across training seeds 42, 100, and 250.

Paired supervisionOptimizerLVBenchVideoMME-Long
NoGMPO35.8±0.550.9±0.6
YesGMPO36.7±0.552.3±0.5
NoTB-GMPO38.4±0.254.2±0.2
YesTB-GMPO39.5±0.256.0±0.2

For a fair comparison, all conditions started from the same replay-verified SFT snapshot before paired-label training. Supervision branches matched trajectory data, continued-SFT updates, and learning rates, differing only in the action-label loss. Within each branch, optimizers shared RL initialization, reference checkpoint, data, rewards, rollout rules and budgets, learning rate, gradient clipping, and seeds.

Both components improve accuracy under either setting of the other component. Paired supervision improves LVBench/VideoMME-Long accuracy by 0.9/1.4 points under GMPO and 1.1/1.8 points under TB-GMPO. TB-GMPO improves accuracy by 2.6/3.3 points without supervision and 2.8/3.7 points with it. Their combination achieves the best accuracy, improving over GMPO without paired supervision by 3.7/5.1 points. The interaction, defined as the difference between the conditional gains, is 0.2/0.4 points. These results support distinct, complementary contributions with approximately additive gains.

2. Training-state coverage, action distribution, and label stability

We attempted paired Monte Carlo evaluation on 6,840 training states. Of these, 6,651 completed all three action evaluations and 4,218 received clear labels. We excluded 2,433 completed states because their action-value differences did not pass the interval-and-margin rule; 189 evaluations were incomplete. Ambiguous states contribute no action-label loss.

Table R2. Training-state coverage and accepted-label distribution.

(a) Training-state coverage

Evaluation stageStates
Candidate7,284
Attempted6,840
Completed6,651
Accepted4,218
Ambiguous2,433
Incomplete189

Acceptance rates: 61.7% of attempted, 63.4% of completed, and 57.9% of candidate states.

(b) Accepted-label distribution

ActionStatesRate (%)
STOP1,77242.0
LOW1,40133.2
HIGH1,04524.8

Each completed state uses 32 continuations per action (96 total) , with λC=0.10 , τV=0.02 , and 1,000 bootstrap resamples . Acceptance requires the label’s lower bound to exceed both alternatives’ upper bounds by τV, using Bonferroni-corrected intervals with nominal 95% simultaneous coverage . Actions share the remaining budget; LOW/HIGH use identical timestamps at different resolutions.

We vary the Monte Carlo budget, value margin, confidence level, and frame density. Table R3 reports jointly confident state counts and label agreement.

Table R3. Fixed-state sensitivity and independent-repeat evaluation. Default: M=32, λC=0.10, τV=0.02, simultaneous confidence 95%. Non-default rows are compared with the default.

SettingCompletedAcceptance (%)Jointly confidentAgreement (%)STOP/LOW/HIGH (%)Confident→ambiguous (%)
Default78463.5——42.2/33.1/24.7—
M=1678058.642092.943.1/33.0/23.911.2
M=6478666.347496.241.8/33.3/24.94.4
τV=0.0178470.848794.041.5/33.5/25.02.8
τV=0.0478452.740497.543.0/32.6/24.418.1
Confidence 90%78467.148094.841.9/33.2/24.93.6
Confidence 99%78458.042697.242.8/32.9/24.313.7
Frame density ×0.5; cap fixed77960.243190.745.0/32.8/22.29.6
Frame density ×2; cap fixed78264.145993.940.9/33.4/25.76.2
Independent repeat; new rollout seeds78362.845194.742.5/32.9/24.67.0

Independent rollouts achieve 94.7% agreement across 451 jointly confident states , with agreement ranging from 92.9–96.2% across Monte Carlo budgets, 94.0–97.5% across value margins, and 90.7–93.9% across frame densities. Accepted labels are generally repeatable, though acceptance depends on the threshold and sampling density. Agreement reflects repeatability, not ground-truth accuracy.

AuthorsOpenReview ↗

Response to Reviewer xjPX (2/2)

3. Baseline input resolution and the official LVBench discrepancy

We agree that the official Qwen2.5-VL-7B-Instruct model card reports 45.3% on LVBench. Our original Qwen baseline used 64 uniform frames at 448×448 and obtained 38.3%, while MindSight achieved 39.5% under its original browsing budget.

3.1 Updated evaluation under comparable higher frame budgets

As shown in Table R4a, with comparable higher frame budgets (up to 488 frames for MindSight versus 512 frames for Qwen), MindSight achieves 48.5%, outperforming the updated local official-style Qwen baseline (44.7%) by 3.8 percentage points.

Compared with the updated fixed-resolution Qwen baseline, MindSight achieves higher accuracy (48.5% vs. 43.8%) with a lower maximum visual-token budget (50,176 vs. 65,536) and a smaller context limit (64K vs. 128K).

Table R4a. LVBench configurations and comparison under comparable higher frame budgets. Original low-budget results are listed separately; visual budgets count maximum acquired visual LLM tokens.

ModelFrame budgetProcessed resolutionMaximum acquired visual tokensContextLVBench(%)
Original Table 1 Qwen backbone64 frames448×4488,19264K38.3
Original Table 1 MindSight-7BScan ≤64; Focus ≤24/attempt; Max. 5 attemptsScan/LOW: 224×224; HIGH: 448×44817,40864K39.5
Qwen2.5-VL-7B, updated single pass512 frames448×44865,536128K43.8
Qwen2.5-VL-7B, updated local official-style setting512 frames; dynamic-fps samplingDynamic pixel rule--128K44.7
MindSight-7B, updated evaluationScan ≤128; Focus ≤72/attempt; Max. 5 attempts; Total ≤488 framesScan/LOW: 224×224; HIGH: 448×44850,17664K48.5

Updated MindSight setting uses 224×224 scan/LOW inputs and 448×448 HIGH inputs, with at most 128 scan frames, 72 frames per focus, at most five focus attempts, and a 64K context. Its maximum acquired visual budget is 50,176 tokens, excluding text and special tokens. The Qwen sweep fixes spatial resolution at 448×448, as well as questions, modalities, prompt, parser, and decoding; the context limit is expanded as needed.

3.2 Qwen frame-budget sweep and local official-style evaluation

Table R4b reports target frame budgets, processing resolutions, and visual-token counts calculated for the full target inputs.

Table R4b. Single-pass frame-budget sweep and local official-style evaluation. The controlled sweep fixes spatial resolution; the final row uses a separate dynamic-fps and dynamic-pixel configuration.

SettingTarget framesProcessed resolutionAcquired visual tokensContextLVBench (%)
Single pass64448×4488,19264K38.3
Single pass128448×44816,38464K40.2
Single pass256448×44832,76864K42.1
Single pass512448×44865,536128K43.8
Local official-style setting512Dynamic pixel rule--128K44.7

At a fixed resolution of 448×448, increasing Qwen’s target frame budget from 64 to 512, with the context limit expanded as needed, raises accuracy from 38.3% to 43.8%, a 5.5-percentage-point gain. We could not reproduce Qwen's official 45.3% locally: the closest local Qwen setting reaches 44.7%, leaving a 0.6-point gap.

With comparable higher frame budgets (MindSight: Total ≤488 frames; Qwen: 512 target frames), MindSight outperforms the stronger local Qwen baseline by 3.8 percentage points (48.5% vs. 44.7%), showing that its advantage persists after increasing the baseline's input budget. Together with the controlled Qwen sweep, these results explain why the original restricted-budget scores were lower and support MindSight's benefit in the updated comparison. Official source: Qwen2.5-VL-7B-Instruct model card.

We will replace the original low-budget main results with these updated higher-budget evaluations. The revised paper will report actual frame counts, resizing rules, and visual-token costs.

Thank you again for the time and care you have devoted to reviewing our work. We truly appreciate your thoughtful feedback, which has helped us clarify and strengthen the paper. We warmly welcome any further questions or suggestions and would be happy to address them.

Reviewer_XyA4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces an active perception agent that uses a scan–focus–reflect strategy to identify candidate temporal intervals and adaptively decide whether to stop or inspect an interval at low or high visual fidelity, balancing the value of new evidence against the cost of observation. The policy is learned in a unified manner from multi-turn browsing trajectories, improving accuracy on long-video benchmarks while reducing computational complexity.

Critical strengths
  • This paper presents an end-to-end learning framework for long-video understanding that can be optimized with a unified loss function, offering a more elegant formulation than existing multi-stage methods.
  • Experimental results show the strong performance on four long-video benchmarks and the improving inference efficiency.
Critical weaknesses
  • Trajectory bootstrapping is used to collect data for supervised fine-tuning (SFT). However, filtering trajectories solely based on the correctness of the final answer may not reliably ensure the quality of the collected data, particularly for long trajectories.

  • The large number of hyperparameters makes the learning process difficult to tune and potentially unstable (for example, lambda_{\alpha}, \lambda_{t}, \lambda_{f} in the loss function).

Questions for authors
  • Please provide more details on the techniques used to stabilize training, particularly the tuning of the hyperparameters in the loss function.

  • The reinforcement learning stage may depend heavily on the quality of the SFT model. If the SFT model is not adequately trained, it may be difficult to sample valid and informative trajectories for reinforcement learning. How is this issue addressed?

3 comments

We sincerely thank the reviewer for recognizing the motivation, clarity, and empirical benefits of MindSight. Below, we address the concerns about trajectory quality, training stability, and SFT initialization. Results are reported over three independent runs with seeds 42, 100, and 250, consistent with the main paper.

1. Quality of the SFT trajectories

We agree that a correct final answer alone does not establish trajectory quality.Therefore, we combine answer filtering, dependency-preserving subsequence extraction, and replay verification. From each successful exploration, we retain a short executable subsequence preserving observation request–evidence dependencies, then replay it to check execution consistency and action–observation correspondence. This yields approximately 3K SFT trajectories from 25K candidate video–question examples, without supervising entire long successful traces.

While replay verifies execution consistency, paired action supervision further evaluates observation choices using 32 continuations per action to estimate the downstream returns of STOP, LOW, and HIGH. LOW and HIGH are compared at identical timestamps and with the same remaining budget; ambiguous comparisons are excluded from the action-label loss.

2. Training stability and reward hyperparameters

As specified in the Experimental Setup, we use (λa,λt,λf)=(0.5,0.2,0.3)(\lambda_a, \lambda_t, \lambda_f) = (0.5, 0.2, 0.3) for answer correctness, temporal grounding, and format validity, respectively. Answer correctness receives the largest individual weight , while the auxiliary terms encourage relevant observations and protocol compliance. Format validity is averaged over assistant turns, avoiding reward accumulation merely from additional turns. Temporal reward uses only successful focus intervals, excluding direct contributions from the global scan.

For stable group-relative optimization, we sample eight trajectories per prompt at temperature 1.0 and top-p = 0.95, while KL regularization limits abrupt drift from the SFT reference . TB-GMPO additionally uses equal turn weights, group-wise return normalization, advantage clipping at 2, a log-ratio guard with c=1.5c=1.5, and pessimistic clipping with ϵ=0.2\epsilon=0.2. Parameter-gradient norms are clipped at 10. Additional training hyperparameters are summarized in Table 4 of the appendix.

We evaluate alternative reward weights to assess sensitivity to the balance between task accuracy and auxiliary rewards. The selected configuration achieves the highest mean accuracy on both benchmarks among those tested, supporting its use in the main experiments:

Reward weights (λ_a, λ_t, λ_f)LVBench accuracy (%)VideoMME-Long accuracy (%)
(0.5, 0.2, 0.3)39.5 ± 0.256.0 ± 0.2
(1.0, 0.0, 0.0)37.8 ± 0.353.7 ± 0.4
(0.6, 0.2, 0.2)38.7 ± 0.355.0 ± 0.3
(0.5, 0.3, 0.2)39.0 ± 0.255.5 ± 0.3
(0.4, 0.4, 0.2)37.9 ± 0.454.2 ± 0.4
(0.4, 0.2, 0.4)38.4 ± 0.354.8 ± 0.3
(0.34, 0.33, 0.33)38.1 ± 0.354.5 ± 0.4

3. Dependence of RL on the quality of the SFT model

We agree that effective RL requires a sufficiently capable initialization. We assess checkpoint readiness through execution validity and reward variation before RL , rather than assuming any SFT checkpoint is suitable.

Starting from an instruction-tuned backbone, replay-verified SFT and paired action supervision establish executable browsing behavior before TB-GMPO. The following table compares the backbone and the main training variants:

Training variantLVBench accuracy (%)VideoMME-Long accuracy (%)
Qwen2.5-VL-7B-Instruct (backbone)38.3 ± 0.150.0 ± 0.1
RL only36.1 ± 0.351.7 ± 0.3
SFT only34.8 ± 0.351.1 ± 0.3
Full (SFT + RL)39.5 ± 0.256.0 ± 0.2

The checkpoint used to initialize RL retains meaningful task competence. It achieves 60.0% accuracy before RL on the held-out RL validation split (Figure 4, appendix), which is separate from the benchmark evaluation sets. Although SFT only accuracy decreases by 3.5 percentage points on LVBench relative to the backbone, it improves VideoMME-Long by 1.1 percentage points.

These results do not suggest that RL can compensate for arbitrarily weak initialization. Rather, we consider it essential to use a readiness assessment to select a sufficiently capable SFT checkpoint before RL . From this initialization, TB-GMPO raises validation accuracy to 72.5% and improves LVBench and VideoMME-Long accuracy by 4.7 and 4.9 percentage points, respectively. Full (SFT + RL) outperforms both the SFT only and RL only.

Thank you again for your careful review. Please feel free to raise any further questions; we will be happy to provide detailed responses.

We would sincerely appreciate it if you could reconsider your evaluation of our work in light of these clarifications.

I remain concerned about the large number of hyperparameters, which may make the proposed method sensitive to their settings and limit its ability to generalize to other domains. I therefore maintain my initial rating.

AuthorsOpenReview ↗

Clarification on Reward-Weight Tuning and Fixed Training Settings

We sincerely thank the reviewer for the continued discussion and the opportunity to clarify our hyperparameter choices.

The tuning discussed in Section 2 of our previous response concerns only the three reward weights (λa,λt,λf)(\lambda_a,\lambda_t,\lambda_f), which balance answer correctness, temporal grounding, and format validity. The other reported settings are standard sampling and training configurations, together with fixed stabilization settings in TB-GMPO. These settings are used as fixed defaults and require no additional tuning in our experimental setup. We report them for reproducibility.

We have already provided a comparison of seven reward-weight configurations, with all remaining settings held fixed and results averaged over seeds 42, 100, and 250. The selected setting (0.5,0.2,0.3)(0.5,0.2,0.3) achieves the highest mean accuracy on both benchmarks among all configurations tested, providing direct experimental support for our choice.

We appreciate the reviewer's attention to the practical burden of tuning. We would be grateful if this clarification and the ablation evidence already provided could be taken into account when reconsidering this concern and the overall evaluation. Thank you again for your time and careful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.