acceptodds
Under review as a conference paper at ICLR 2027

Reasoning Beyond Imitation: Observation Conditioned Distillation for Video LLMs

10 views

Abstract

Reasoning distillation offers a promising approach to transfer temporal reasoning capabilities of powerful Video-LLMs to open-source student models. However, teacher and student Video-LLMs may construct different temporal observations from the same video due to the adoption of distinct video sampling pipelines. A reasoning trajectory valid under the teacher's observations may rely on evidence inaccessible to the student and thus cannot be directly executed based on the student's observations. We formalize this as a conditional target mismatch, thereby revealing the Reasoning Acquisition–Execution Gap. To bridge this gap, we propose an acquisition-to-execution framework that first distills teacher trajectories into a student reasoning prior and then further optimizes the execution capability of this prior across multiple observations. We further introduce Execution-Aware Reference Regularization, which retains the existing prior when it is executable under current observations, while allowing a greater degree of adaptive adjustment when it is not. Experiments across multiple Video-LLM backbones and six video reasoning benchmarks show that explicitly optimizing observation-conditioned execution consistently outperforms supervised distillation and standard RL-based optimization, with gains extending to held-out out-of-domain benchmarks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Video LLMs, Knowledge Distillation, Video Reasoning, Reinforcement Learning

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_4EqYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates Teacher–Student Observation Mismatch in video reasoning distillation and proposes a framework combining SFT, Multi-Observation GRPO (MO-GRPO), and Execution-Aware Reference Regularization (EARR).

Critical strengths

The paper presents a clear and well-organized experimental evaluation across multiple Video-LLM backbones and six benchmarks. The observation-sensitivity analysis also highlights a relevant challenge in video reasoning.

Critical weaknesses

Motivation–method mismatch. The paper argues that teacher and student models observe different video frames, but MO-GRPO only optimizes student-generated trajectories under different sampling configurations. It is unclear how this directly addresses the proposed teacher–student mismatch. The SFT part includes, but from the ablation, it does not perform well (performance degradation on Qwen3-VL).

Insufficient analysis of MO-GRPO. The paper does not convincingly explain why multi-observation training improves reasoning beyond ordinary temporal data augmentation.

Limited ablation evidence. MO-GRPO does not consistently outperform standard GRPO. Moreover, the missing GRPO + EARR baseline makes it difficult to isolate the contribution of MO-GRPO.

Questions for authors
  1. Can the authors experimentally validate the impact of matched versus mismatched teacher–student observations?
  2. Why does MO-GRPO improve performance beyond standard temporal data augmentation?
  3. How does standard GRPO + EARR perform compared with the full method?
4 comments

Thank you for your careful review. We answer each question below and have uploaded a revised PDF with the changes marked in blue. Table and section numbers refer to the revised PDF.

Weakness 1 / Question 1: Motivation–method mismatch; matched vs. mismatched validation

We ran the matched-versus-mismatched experiment. Teacher and student are from the same family (Qwen3-VL-235B-A22B-Thinking and Qwen3-VL-4B-Thinking), so the observation is the only factor that varies; the student always receives 2 FPS, φ = 0. When the teacher receives the same frames, SFT keeps the student's accuracy on TimeBlind (64.50 → 65.43). When only the teacher's phase is shifted to φ = 0.5, with everything else identical, SFT lowers the student to 53.88. The mismatch alone accounts for an 11.55-point difference.

This also clarifies how the method addresses the mismatch. It does not try to align the student with the teacher's frames, because a black-box teacher does not expose them: we send the video, and which frames it reads is internal. Instead, Stage II removes the student's dependence on the teacher's observation. The student generates trajectories from its own observation and is rewarded by the task outcome, so the mismatched target is no longer in the objective. We use multiple observations for the same reason: since the teacher's frames are unknown, we cannot know where a given student observation departs from them, so the student is trained to execute correctly under several observations rather than one.

The results match this account. In the black-box setting, SFT leaves the student below the pretrained model under all five observations in Table 2 (mean 56.80 vs. 62.03), and the full method brings it above the pretrained model under all five (mean 65.65).


Weakness1 (cont.): Why keep Stage-I SFT when it degrades Qwen3-VL?

Stage I is not kept for its standalone accuracy. It supplies the prior that Stage II turns into accuracy, and Stage II does not work without it: multi-observation RL from the pretrained model reaches 51.98, below the pretrained model itself (54.38), whereas the same RL from the SFT checkpoint reaches 54.35 and is higher on all three benchmarks, and 56.31 with EARR (Fig. 5).

The drop after SFT is also not a failure of SFT. Our controlled experiment shows that it comes from the observation mismatch: with identical teacher and student frames SFT keeps the student's accuracy (64.50 → 65.43 on TimeBlind), and with shifted teacher frames it falls to 53.88. A black-box teacher does not let us remove this mismatch, so SFT alone pays its cost, and that is exactly why Stage II follows it. The degradation is the observation the paper starts from, not a flaw in the pipeline.

Weakness 2 / Question 2: Why does MO-GRPO improve beyond ordinary temporal augmentation?

On its own it does not, and this is the point. Training on several views under one uniform constraint is temporal augmentation in the RL setting, and it brings no gain over single-view GRPO: 54.35 against 55.16 on the Fig. 5 average, and 50.69 and 52.52 at the two other constraint strengths we tested (β = 0.001 and 0.005).

The gain comes from what augmentation does not do: treating the views of the same question differently. Augmentation regards views as interchangeable samples. Our method regards each view as a separate execution condition and sets the constraint to the SFT prior per view, according to whether that prior answers correctly under it. This matters because the prior is acquired under one sampling configuration and is less reliable under others (in Table 2 the SFT model scores 53.75 and 52.50 at 1 FPS, against 60.12 at 2 FPS, φ = 0), so a uniform constraint ties the policy to the prior even where it fails. With per-view control the average rises to 56.31; assigning the same two coefficients to views at random brings it back to 54.26. Multi-observation training is what exposes these per-view differences, and EARR is what uses them.


Weakness 3 / Question 3: MO-GRPO does not consistently outperform standard GRPO. How does standard GRPO + EARR perform compared with the full method?

GRPO + EARR is below the full method on all three benchmarks: 64.38 vs. 66.50 on TimeBlind, 66.25 vs. 67.47 on MVBench and 34.02 vs. 34.97 on TOMATO, or 54.88 vs. 56.31 on average. The full ablation, including the new controls, is given below (accuracy in %, Qwen3-VL-4B-Thinking; all Stage-II variants start from the same Stage-I SFT checkpoint).

Stage IIReference KLTimeBlindMVBenchTOMATOAvg.
SFT-60.1262.0024.3048.81
+ GRPOconst 0.00365.8766.0033.6255.16
+ GRPOEARR64.3866.2534.0254.88
+ MO-GRPOconst 0.00166.6265.2520.2150.69
+ MO-GRPOconst 0.00362.8866.1234.0654.35
+ MO-GRPOconst 0.00562.5064.7530.3252.52
+ MO-GRPOshuffled EARR65.3866.1331.2654.26
+ MO-GRPOEARR (ours)66.5067.4734.9756.31

This table isolates the contribution of each component. We agree that MO-GRPO alone does not outperform GRPO (54.35 vs. 55.16), and EARR alone does not either (54.88 vs. 55.16). The contribution is their combination: adding multi-observation training to EARR raises the average from 54.88 to 56.31, and adding EARR to multi-observation training raises it from 54.35 to 56.31. EARR sets the constraint per view, so it needs the several views of each question that only multi-observation training provides (see W2/Q2).

The remaining rows rule out the alternative explanations. The gain is not a matter of constraint strength: no constant coefficient reaches the full method's average (50.69, 54.35 and 52.52 for β = 0.001, 0.003 and 0.005), and EARR's mean coefficient in the single-observation run is 0.00325, close to the constant 0.003. Nor is it a matter of mixing two strengths: assigning EARR's two coefficients at random gives 54.26, on par with the constant baseline. Only matching the coefficient to the reference outcome under each view produces the gain.

AuthorsOpenReview ↗

Looking forward to your feedback

Dear Reviewer 4EqY,

Thank you again for your review. We posted our response on 13 Nov, with a new experiment for each of your three questions.

Could you let us know whether it addresses your concerns, in particular on soundness? If anything remains open, we would like to resolve it while the discussion period is still running. If your concerns are addressed, we would appreciate it if you could update your assessment.

Best regards,

The Authors

AuthorsOpenReview ↗

Kind Follow-up on Our Rebuttal and Additional Experiments

Dear Reviewer 4EqY,

We would like to kindly follow up on our previous rebuttal, where we provided additional experiments addressing all three of your questions, particularly the concerns regarding the soundness of our method.

As the discussion period is approaching its end, we would sincerely appreciate your feedback on whether these additional results address your concerns. If so, we would be grateful if you could reconsider your assessment.

Thank you very much for your time and consideration.

Best regards,

The Authors

Reviewer_2mvvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that in the current video understanding model distillation methods, when the closed-source good model, the teacher and the student sample different frames from the same video, you cannot determine the sampling frames of the teacher model. Using such teacher reasoning trajectory distillation to generate student trajectories is of no use. The authors refer to this failure phenomenon as "Reasoning Acquisition–Execution Gap". The authors made two contributions: the first stage imitates the correct teacher trajectory under the student observation condition; the second stage runs GRPO on five student sampling configurations and sets the KL weight in EARR to 0.005 or 0.001 based on whether the first-stage frozen model has correctly answered. The teacher model is Gemini 2.5 Pro, and the student models are 4B and 8B-parameter Qwen3-VL and InternVL3.5. Training uses TimeBlind, MVBench, and TempCompass datasets, and TOMATO, Video-MME (short), and VideoReasonBench are used for testing. The paper explains this phenomenon, which is very reasonable and reliable, but the authors' proposed methods and experiments have not proved that the problem has been solved.

Critical strengths

The discovered phenomenon: Different time samples of the same video, especially the sampling of the closed-source model, are not as clear, will cause changes in the correctness of the video large language model (Video-LLM), and the current model distillation method does not share the same sampling method.. Evidence: Fig. 2 keeps the frame rate fixed and still flips answers when only the phase changes; 37.9%–58.4% of TimeBlind questions are observation-dependent. Table 3 shows that Gemini, Qwen3-VL, and InternVL3.5 select frames by different rules. Why it matters: This phenomenon has been confirmed, which is also the reason why the scoring result is a weak rejection rather than a clear rejection.

Critical weaknesses

1 The method does not treat the phenomenon it defines. E.g. Sec. 2 defines the gap as teacher evidence that the student cannot see. The paper never measures teacher–student frame overlap. Stage I still imitates the teacher trajectory on the student observation (Eq. 1). Stage II discards that trajectory (Appendix C.3, D.2) and rewards only the final answer and the output format. EARR does not look at frames; it sets β to 0.005 if the frozen Stage-I answer is correct and to 0.001 otherwise. 2 The experiments do not show that the pipeline solves it. Fig. 5 is the direct ablation on Qwen3-VL-4B. The average is 54.38 for the pretrained baseline, 48.81 after Stage-I SFT, 51.98 for multi-observation GRPO without SFT, and 54.35 for the two stages together. Only adding EARR reaches 56.31.

Questions for authors

1 On the training set, how often does the teacher trajectory cite a frame or event that is absent from the student observation? And does SFT recover if the teacher is run with the student’s sampler? — If missing evidence is rare, or if matched-sampling SFT removes the drop, the named phenomenon is not what this method addresses, and I would keep Rating at 2 or move it to 1. If missing evidence is common and the two-stage model, without the KL switch, beats the pretrained baseline on those items under a matched rollout budget, I would reconsider Rating at 3.

2 The proof of the phenomenon needs to be more comprehensive. Currently, it is defined through a formal problem statement that is too simplistic, and some details are not clearly explained. However, the problem itself is meaningful, and we hope it can be elaborated better.

5 comments

Thank you for your careful review. We answer each question below and have uploaded a revised PDF with the changes marked in blue. Table and section numbers refer to the revised PDF.

Question 1 (a): How often does the teacher trajectory cite a frame or event that is absent from the student observation?

We measured this at the level of cited evidence. On the 800 held-out TimeBlind questions, each model answers from its own observation (the teacher, Gemini 2.5 Pro, through its native video interface; the student, Qwen3-VL-8B-Thinking, from its sampled frames), and we compare the evidence they cite on the 467 questions both answer correctly.

The restriction to jointly correct questions is deliberate. When one model is wrong, a difference in evidence may only reflect its error. When both are right, a difference means the teacher's evidence is not the evidence the student relied on under its own observation.

Protocol. Claude Sonnet 4.6 compared the two responses for each question, anonymized and in randomized order, and separated semantic differences in evidence from paraphrasing.

Evidence relationshipQuestionsShare
Same evidence30665.5%
Same key evidence, different details10823.1%
Different key evidence5311.3%

In 161 questions (34.5%) the two cite different evidence. In 53 of them (11.3%) the difference is in the key evidence: 40 use a different answer-bearing evidence chain and 13 make incompatible key factual claims.

What this means for distillation. Stage I keeps a teacher trajectory whenever its answer is correct. The audit shows that a correct answer does not make the trajectory a target the student can follow: in one in three such cases the cited evidence differs from what the student uses, and in one in nine the key evidence itself differs. The controlled experiment in Q1 (b) shows what such targets do to SFT.


Question 1 (b): Does SFT recover if the teacher is run with the student's sampler?

Yes, and this is what the paper's claim predicts: if the observation mismatch is what breaks SFT, removing the mismatch must remove the drop. With the open-source teacher, whose sampler we can set, SFT under identical teacher/student frames keeps the student's accuracy (64.50 → 65.43 on TimeBlind), while shifting only the teacher's sampling phase lowers it to 53.88 (details in our response to Reviewer qpsm, W2/Q2).

This result confirms the phenomenon; it does not replace the method. The problem we address is distillation when the mismatch cannot be removed:

  • Our setting is a black-box teacher (§1). We send it the video; which frames it reads, and at what phase, is internal. Matched sampling is not an option there.
  • Setting the nominal FPS does not fix this. In Fig. 6, two APIs that both sample at 1 FPS return disjoint frames from the same clip (15, 45, 75, … vs. 1, 31, 61, …). The controlled experiment shows that such a phase shift alone costs 10.62 points.

The method is built for exactly this case. Stage II never needs the teacher's frames: it optimizes the student under its own observations, and in the black-box setting it raises the SFT student from 60.12 to 66.50 on TimeBlind.

Weakness 2 / Question 1 (c): Can the two-stage model without the KL switch (EARR) outperform the pretrained baseline under a matched rollout budget?

Yes, and with a smaller budget than the full method. The KL switch is then what makes multi-observation training pay off, and it is the observation-conditioned part of the method, not an add-on.

  1. Without the switch, the two-stage model already beats the pretrained model. SFT + GRPO with a constant KL uses four rollouts per question, one-fifth of the full method's budget, and averages 55.16 against 54.38 for the pretrained model (Fig. 5). Across four backbones and six benchmarks it is above the pretrained model in 17 of 24 settings (Table 1). The 54.35 quoted in the review is the multi-observation variant under a uniform constraint, which is the configuration the switch is designed to replace.

  2. The switch is the observation-conditioned component. All variants below use the same rollouts (five observations × four rollouts); EARR adds one deterministic reference decode per observation, which is not used as a rollout.

Reference KL (with MO-GRPO)TimeBlindMVBenchTOMATOAvg.
const 0.00166.6265.2520.2150.69
const 0.00362.8866.1234.0654.35
const 0.00562.5064.7530.3252.52
EARR coefficients assigned at random65.3866.1331.2654.26
EARR66.5067.4734.9756.31

No constant strength reaches EARR's average, and assigning the same two coefficients at random gives 54.26, on par with constant 0.003. The gain appears only when the coefficient is tied to whether the SFT prior succeeds under the current observation, which is the executable / non-executable distinction of §2 applied per observation.

  1. With the switch, the gain is consistent. The full method is above the pretrained model in 23 of 24 settings in Table 1 (the exception is Qwen3-VL-4B on Video-MME, 76.00 vs. 76.67) and above SFT + GRPO in 23 of 24 (the exception is Qwen3-VL-8B on Video-MME, 80.22 vs. 81.04).

Weakness 1: The method does not treat the phenomenon it defines

The method treats the phenomenon by removing the student's dependence on the teacher's observation. It does not inspect the teacher's frames, because a black-box teacher does not expose them. Stage I is acquisition, and §4.1 already states that its targets are mismatched; we keep it because the prior it yields is needed (Stage II reaches 51.98 without it and 54.35 with it, Fig. 5). Stage II then discards the teacher trajectory, and this is the treatment itself: the mismatched target leaves the objective, the student generates trajectories from its own observation, and the reward is the task outcome, which is valid under any observation.

EARR does not look at frames because it uses the observable symptom of the mismatch instead: whether the SFT prior succeeds under the current student observation. Our controlled experiment shows that mismatch is what makes this prior fail (65.43 with identical teacher/student frames, 53.88 with shifted ones), and assigning EARR's two coefficients at random removes its gain (54.26 vs. 56.31).

On frame overlap: where the teacher's sampler can be set, 100% vs. 0% shared frames gives the 65.43 vs. 53.88 above (Q1 (b)); for the black-box teacher, Q1 (a) measures the mismatch at the evidence level.

Question 2: A more complete statement of the phenomenon;

We appreciate that you find the problem meaningful. We have rewritten Sec 2 so that the phenomenon is stated precisely and then verified step by step.

Formally, the teacher writes its trajectory from its own observation, yT∼pT(⋅∣OT,q)y_T\sim p_T(\cdot\mid O_T,q), where pTp_T denotes the retained teacher trajectories, but the student is trained to produce it from OSO_S. The SFT optimum is therefore πSFT∗(y∣OS,q)=EOT∣OS,q[pT(y∣OT,q)]\pi^*_{\mathrm{SFT}}(y\mid O_S,q)=\mathbb{E}_{O_T\mid O_S,q}[p_T(y\mid O_T,q)].

an average over teacher observations the student never receives. The target a student can actually execute is pT(y∣OS,q)p_T(y\mid O_S,q), the teacher's reasoning on the student's own frames. We define the conditional target mismatch as the divergence between the two. It is zero when OT=OSO_T=O_S and positive whenever the teacher's trajectory depends on frames the student does not have; in that case SFT trains the student to state evidence it cannot check.

The verification follows the same chain. Observations do differ in practice: in Fig. 6, two APIs that both sample at 1 FPS return disjoint frames from the same clip. The difference changes the evidence: on questions both models answer correctly, teacher and student cite different evidence in 34.5% of cases and different key evidence in 11.3% (Q1 (a)). And the difference breaks SFT: with OT=OSO_T=O_S the student keeps its accuracy (64.50 → 65.43 on TimeBlind), while shifting only the teacher's frames lowers it to 53.88 (Q1 (b)).

Reviewer_qpsmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Temporal sampling can give teacher and student different evidence, making teacher rationales unsuitable for the student's input. The paper uses SFT followed by multi-observation GRPO and Execution-Aware Reference Regularization (EARR), which tightens KL when the SFT reference is correct for the current observation and relaxes it otherwise. Across two model families, two sizes, and six benchmarks, the full pipeline usually improves accuracy, but the ablations do not identify which component drives the gains.

Critical strengths
  1. The full method improves nearly all backbone-benchmark combinations, including held-out benchmarks, against SFT+GRPO, LUFFY, and T-GRPO. Shared initialization and the same RL subset make the comparison informative.

  2. EARR conditions regularization on reference correctness, preserving useful behavior while adapting on failures. Observation-local advantages also avoid comparing trajectories from different frames.

Critical weaknesses
  1. The ablations do not separate multi-observation optimization from EARR or KL strength. On TimeBlind, SFT+MO-GRPO (62.88) is below SFT+GRPO (65.87), while the full system improves after adding EARR. Missing single-observation GRPO+EARR, tuned constant-KL, and shuffled-EARR controls leave open whether adaptive regularization, rather than multi-observation training, explains most of the gain.

  2. The evidence supports sensitivity to sampling, but not the proposed acquisition-execution explanation. Changing observations changes correctness, yet does not quantify teacher-student mismatch; the Qwen3-VL-Thinking degradation may instead reflect disrupted reasoning. Qualitative examples and answer rewards do not show that rationales use sampled frames. Compare identical versus deliberately mismatched teacher/student frames under matched settings; otherwise the paper demonstrates useful post-training without isolating its stated mechanism.

Questions for authors
  1. How does single-observation GRPO+EARR compare with the full method, tuned constant-KL baselines, and shuffled EARR assignments with the same high/low-coefficient ratio?
  2. With the open-source teacher, compare SFT and RL when teacher and student receive identical versus deliberately mismatched frames, holding other choices fixed. Does the advantage grow with measured evidence mismatch?
5 comments

Thank you for your careful review. We answer each question below and have uploaded a revised PDF with the changes marked in blue. Table and section numbers refer to the revised PDF.

Weakness 1 / Question1: Separating multi-observation training, EARR and KL strength (GRPO+EARR, tuned constant KL, shuffled EARR)

We ran all three requested controls. None reproduces the full method: EARR without multi-observation training gives no gain, no constant coefficient reaches the full method's average, and assigning EARR's two coefficients at random removes its gain.

Accuracy (%) on Qwen3-VL-4B-Thinking. All Stage-II variants start from the same Stage-I SFT checkpoint; Avg. is the unweighted mean of the three benchmarks.

Stage IIReference KLTimeBlindMVBenchTOMATOAvg.
SFT-60.1262.0024.3048.81
+ GRPOconst 0.00365.8766.0033.6255.16
+ GRPOEARR64.3866.2534.0254.88
+ MO-GRPOconst 0.00166.6265.2520.2150.69
+ MO-GRPOconst 0.00362.8866.1234.0654.35
+ MO-GRPOconst 0.00562.5064.7530.3252.52
+ MO-GRPOshuffled EARR65.3866.1331.2654.26
+ MO-GRPOEARR (ours)66.5067.4734.9756.31
  1. EARR without multi-observation training. Single-observation GRPO + EARR averages 54.88, against 55.16 for constant-KL GRPO and 56.31 for the full method. In this run EARR assigns the high coefficient to 56.25% of observations, a mean β of 0.00325, close to the constant 0.003. Adaptive regularization alone therefore does not explain the gain.

  2. Tuned constant KL. With multi-observation training, constant β = 0.001, 0.003 and 0.005 average 50.69, 54.35 and 52.52, all below 56.31. The weakest constraint gives the best TimeBlind score (66.62, 0.12 above the full method) but collapses on the held-out TOMATO (20.21, below the SFT checkpoint's 24.30). β = 0.005 recovers less from the SFT checkpoint than β = 0.003 on all three benchmarks.

  3. Shuffled EARR. Keeping the two coefficients and their high/low ratio but assigning them at random gives 54.26, on par with constant β = 0.003 (54.35) and below the full method on all three benchmarks (by 1.12, 1.34 and 3.71 points). The gain therefore depends on matching the coefficient to the reference outcome, not on mixing two constraint strengths.

Stage I is trained under a single sampling configuration (2 FPS), and single-observation GRPO trains under that same configuration; there the adaptive coefficient brings no benefit. Multi-observation training adds 1-FPS and phase-shifted views, on which the Stage-I model is less reliable: in Table 2 it scores 53.75 and 52.50 at 1 FPS, against 60.12, 58.62 and 59.00 at 2 FPS. A constant coefficient ties the policy to the reference on these views as well. EARR relaxes the constraint on views where the reference fails and keeps it where the reference succeeds.

Fig. 5 includes the single-observation EARR result, and we have added this discussion to §5.4.

Weakness 2 / Question 2: Identical vs. deliberately mismatched teacher/student frames with the open-source teacher

We ran this comparison. With identical teacher and student frames, SFT does not degrade the student (64.50 → 65.43 on TimeBlind). Shifting only the teacher's sampling phase lowers it to 53.88.

Design. To make the observation the only factor that varies, teacher and student come from the same family: Qwen3-VL-235B-A22B-Thinking and Qwen3-VL-4B-Thinking. They share the thinking format and visual preprocessing and differ only in size. The student always receives 2 FPS, φ = 0. The teacher receives either the same frames or frames at 2 FPS, φ = 0.5 (a 0.25 s offset: disjoint frame sets, same frame count). Questions, prompts, teacher generation settings and SFT hyperparameters are identical in both runs. This control is possible because the open-source teacher's sampler can be set by us; the black-box teacher in the main experiments receives the raw video.

Teacher framesStudent framesShared framesTimeBlind after SFTΔ vs. pretrained
(pretrained, no SFT)2 FPS, φ = 0–64.50–
2 FPS, φ = 0 (identical)2 FPS, φ = 0100%65.43+0.93
2 FPS, φ = 0.5 (shifted)2 FPS, φ = 00%53.88−10.62

What this shows.

  1. The observation mismatch is sufficient to cause SFT degradation. The two runs differ only in the teacher's frames and end 11.55 points apart.
  2. Imitating teacher trajectories is safe when the student sees what the teacher saw: under identical frames the student keeps its accuracy. A large gain is not expected here, since the same-family teacher scores 69.75 on TimeBlind, only 5.25 points above the pretrained student.
  3. The drop is not a side effect of the reasoning format, which is the same in both runs (see the next response).

RL under mismatch. In the main setting the black-box teacher receives the raw video, so its frames cannot be matched. There Stage II raises the SFT student from 60.12 to 66.50 on TimeBlind.


Weakness 2 (continue): Could the Qwen3-VL-Thinking drop reflect disrupted native reasoning?

The evidence points to the observation mismatch, not to disrupted native reasoning. Three results:

  1. SFT does not disrupt a Thinking student when the frames match. In the controlled experiment above, teacher and student are both Qwen3-VL-Thinking models, and the student's thinking trace is fine-tuned on teacher trajectories in both runs. With identical frames the student keeps its accuracy (64.50 → 65.43).
  2. With the reasoning format held fixed, the mismatch alone produces the drop. Shifting only the teacher's frames lowers the same student to 53.88. This 10.62-point loss is larger than the 4.38-point SFT drop on TimeBlind in the main setting (64.50 → 60.12), so the mismatch is sufficient to account for a drop of that size.
  3. The drop also appears without a thinking mode. InternVL3.5 is run without an explicit thinking mode (Appendix D.1), and SFT still lowers InternVL3.5-8B on four of six benchmarks: TimeBlind 64.38 → 63.38, MVBench 62.50 → 58.63, TempCompass 77.47 → 76.37, Video-MME 73.43 → 68.11.

Weakness 2 (continue): Does this mismatch causally explain SFT degradation, and can the proposed method mitigate its effects?

Yes on both counts.

Cause. In the controlled experiment above, the two SFT runs differ only in the teacher's frames: 65.43 with identical frames, 53.88 with shifted frames.

Mitigation. In the main setting the black-box teacher's frames cannot be matched, and SFT lowers Qwen3-VL-4B-Thinking on all six benchmarks. Stage II then optimizes the student on its own observations and recovers the loss: on TimeBlind, 60.12 → 66.50 (pretrained: 64.50), and the full method is above the pretrained model under all five observations in Table 2. The gain depends on relaxing the constraint where the SFT prior fails: assigning EARR's coefficients at random removes it (54.26 vs. 56.31, see W1/Q1).

AuthorsOpenReview ↗

Looking forward to your feedback

Dear Reviewer qpsm,

Thank you again for your review. We posted our response on 13 Nov, including the three controls you requested (GRPO + EARR, tuned constant KL, shuffled EARR) and the identical-vs-mismatched frame experiment.

Could you let us know whether it addresses your concerns? If anything remains open, we would like to resolve it while the discussion period is still running. If your concerns are addressed, we would appreciate it if you could update your rating.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.