acceptodds
Withdrawn from ICLR 2027

ChapterNav: Learning Video Chaptering through Verifiable Navigation Feedback

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automatic video chaptering organizes long instructional videos into reusable directories for efficient content access. Current models typically learn from annotated directories through supervised fine-tuning (SFT) or reference-based optimization. However, a reference directory represents only one plausible organization of a video, and optimizing agreement with it can favor annotation-specific predictions over alternative but equally valid chapter structures. More importantly, higher reference agreement does not necessarily imply greater downstream utility: a directory closer to the ground truth may not be more effective for helping users locate the content they need. We introduce ChapterNav, a framework for evaluating and optimizing video chaptering through verifiable navigation utility. Since chaptering admits multiple valid outputs and lacks a directly verifiable target sequence, ChapterNav instead measures whether a generated directory enables a fixed navigator to locate content satisfying a given information need. Given a query and chapter titles, the navigator selects chapters, while a verifier checks whether the retrieved intervals contain the target content within a prescribed viewing budget. We build navigation benchmarks on two datasets and introduce Successful Navigation Gain (SNG) to quantify navigation utility. The resulting feedback, combined with temporal alignment rewards, is further used for reinforcement learning with verifiable rewards (RLVR). In a paired human study on 50 videos, SNG agrees with human navigation preferences on 76% of pairs, compared with 68% for reference-based evaluation. RLVR achieves a relative improvement of 3.8% on the SNG metric compared with SFT, and a relative improvement of 5.2% compared with reference-reward. These results demonstrate the effectiveness of evaluating and learning video chaptering through downstream content-access utility.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-Video Chaptering, Multimodal Video Understanding, Reinforcement learning with Verifiable Rewards

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QS1GOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ChapterNav to evaluate and optimize video chaptering through navigation utility. A model generates a directory without access to the query, and a frozen text encoder selects chapters based on query–title similarity. The proposed SNG metric combines target-interval coverage with excess viewing cost and supplies feedback for GRPO post-training. The main contributions are the navigation evaluation protocol, query–episode annotations, and navigation-based reward design.

Critical strengths

The motivation is reasonable: agreement with reference chapters does not necessarily imply effective access to desired content. Navigation utility provides a meaningful complementary evaluation perspective. Keeping queries out of the generator input preserves the directory’s role as a reusable navigation structure. Combining coverage and excess access cost also avoids rewarding a trivial whole-video chapter.

Critical weaknesses

Temporal annotation reliability limits the claim of verifiable navigation. Table 6 states that KZKT target timestamps are estimated from text length without per-episode verification against the video clock. Since both SNG and the training reward depend on temporal overlap and a coverage threshold of 0.9, annotation errors can directly change success judgments. The paper does not quantify their effect on system rankings or training gains. There are substantive inconsistencies in the experimental setup and main results. The main text and appendix report 105 versus 70 AVLecture test videos, 436 versus 337 KZKT test videos, and 150 versus 300 training steps. More importantly, Table 2 labels the AVLecture navigator as MiniLM, but its RL navigation results match the Qwen results in Table 3. These discrepancies make it difficult to establish whether the main comparisons use a consistent evaluation setup.

Questions for authors

How large are the discrepancies between estimated KZKT timestamps and the actual video clock? On a manually verified subset, do system rankings and the RL gains over SFT remain consistent? Please specify the data splits, training steps, and navigator actually used for the main results, and explain why Table 2’s MiniLM-labeled results match the Qwen values in Table 3. Do the main conclusions hold under a consistent setup? Do improvements in SNG translate into higher success rates or lower viewing costs when humans actually locate content? What are the actual annotator count, valid sample count, inter-annotator agreement, and statistical uncertainty of the existing preference study?

Reviewer_TpPtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ChapterNav, a framework for evaluating and training video chaptering models through navigation utility. A fixed navigator selects chapters for a query, and interval-based verification measures target coverage and excess viewing. The authors introduce navigation benchmarks and the SNG metric, then use navigation rewards for GRPO post-training. Experiments compare reference-based and navigation-based evaluation, reward designs, and agreement with human directory preferences.

Critical strengths
  • Evaluating chapter directories by their usefulness for content access is well motivated. The distinction between reference agreement and navigation utility is relevant to how chaptering systems are used.
  • SNG combines target coverage with excess viewing cost, avoiding trivial solutions such as selecting the whole video. The reward comparisons and evaluation with a second navigator help assess the proposed approach.
Critical weaknesses
  1. The navigation gains over existing RL rewards are modest, and their consistency is not sufficiently established. The method also requires additional query–episode annotations, human review, and reward computation, but these costs are not quantified. It remains unclear whether the gains are robust and practically meaningful enough to justify the added effort.
  2. The human study provides limited support for the claimed navigation utility. The 76% versus 68% agreement comes from a small preference study without a clear effective sample size or paired significance analysis. It evaluates metric agreement rather than whether RL-generated directories improve users’ navigation outcomes.
Questions for authors
  1. Based on the existing runs, how consistent are the gains over GRACE and the SNG-only reward? Please report any available variation across seeds and summarize the annotation and training costs, so the practical benefit can be assessed. Additional training runs are not requested.
  2. After excluding ties, what are the effective sample sizes for the reported 76% and 68% agreement, and are both computed on the same comparisons? Please use the existing pairwise judgments to assess whether the difference is statistically reliable.

Minor suggestions (no rebuttal response needed):

  • Please replace the template title in the PDF with the paper’s actual title.
  • Please resolve the missing citations and cross-references shown as “?” or “??”, including those in the appendix.
  • Please enlarge the text within the figures to a size closer to the body text and reduce overcrowding.
Reviewer_8HS2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper claims that agreement with a reference chapter partition does not reflect how well a chapter directory supports content access. The authors propose ChapterNav, where a frozen text-embedding navigator ranks chapter titles against a query, and success requires the top-ranked chapter to cover at least 90% of an annotated target episode. SNG combines this success with normalized excess viewing cost. The authors annotate query-episode pairs on AVLecture and a Chinese K-12 collection using LLMs with human review, and use SNG as a GRPO reward on Chapter-Llama SFT checkpoints.

Critical strengths
  1. The problem framing is useful. Treating a chapter directory as a query-agnostic navigation structure, and evaluating it by downstream access rather than reference overlap, is a reasonable direction for an open-ended generation task.
  2. The reward comparison in Tab. 3 is controlled. all RL variants share the same SFT initialization and budget and differ only in the reward, and results are averaged over three seeds without checkpoint selection.
Critical weaknesses
  1. The submission doesn’t appear ready for a complete submission. It contains placeholder values and broken references (like "XX", Section ??). The paper title is even the template title.
  2. In Table 7, R@1 exceeds CondTR for SFT and Reference-RL. This is impossible since CondTR = R@1/BestTR;
  3. The relative gains in the abstract (3.8%, 5.2%) cannot be traced to any table.
Questions for authors
  1. Which navigator was used for each row of Tables 2, 3, and 7? Please report all methods under one navigator.
  2. Which navigator produced the KZKT training reward? Is it the same as the evaluation navigator?

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

Related papers

Discussion (0)

Sign in to comment.