Post-training has become a key stage for eliciting reasoning capabilities in Video Large Language Models (Video-LLMs). Existing post-training methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, they do not fully exploit visual evidence for effective training. RL supervises the policy with a single trajectory-level reward, which is too coarse-grained for video reasoning: the same response-level advantage is applied to every generated token, disregarding that the tokens of a video response are grounded in the visual stream to very different degrees. OPD provides token-level supervision through output-distribution matching, but lacks direct supervision of the spatiotemporal representations that produce those outputs. We observe across extensive experiments that the same model reasons more accurately when its video view combines global temporal context with question-relevant details. Comparing the same student-generated response under this view and uniform sampling reveals changes in token likelihoods and internal representations, providing a training signal for evidence use. Building on this observation, we propose OWLET, an evidence-privileged RL method that derives training signals from a richer teacher view. This view couples uniform global context with question-conditioned local windows, providing visual evidence without requiring a larger model. To better exploit this privileged visual evidence, OWLET introduces (i) evidence-aware credit redistribution, which compares token likelihoods across the two views and redistributes positive response-level advantages according to their relative evidence support; and (ii) spatiotemporal state alignment, which aligns the two views' hidden states on verifier-accepted responses.
Figures & tables
Figure 1: Illustration of how OWLET improves video reasoning through evidence-guided post-training. Uniform frame sampling misses the temporal transitions needed to answer the question, whereas OWLET learns from a correct evidence-privileged teacher.
Figure 2: (a, b) Evidence-only views improve overall accuracy but lose global-temporal accuracy; only pairing global context with local windows improves both. (c) Half of the teacher–student log-probability gap concentrates on 8.5% of response tokens, so spreading training credit uniformly dilutes the evidence signal. (d) Output-only distillation collapses the output KL, yet the hidden divergence plateaus: output supervision alone does not effectively align spatiotemporal representations.
Figure 3: Overview of OWLET . The student samples responses from uniformly sampled frames, while the teacher scores the same responses using a view that combines global temporal coverage with question-conditioned local windows (red box). OWLET consists of two components: (1) Evidence-aware credit redistribution uses teacher–student token log-probability contrasts to redistribute positive advantages within each response; non-positive advantages remain unchanged; (2) Spatiotemporal state alignment aligns hidden states through cosine similarity on verifier-accepted responses.
Backbone
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Qwen3.5-0.8B
Vanilla
49.4
51.8
48.9
15.8
41.9
41.6
SFT
49.5
52.2
49.5
15.7
42.1
41.8
GRPO
50.8
53.9
50.4
18.5
43.3
43.4
OPSD
50.9
53.9
51.4
17.2
44.1
43.5
OWLET
54.9
58.6
55.9
19.9
45.2
46.9
Qwen3.5-4B
Vanilla
64.7
67.4
58.0
24.7
50.3
53.0
Table 1: Accuracy (%) results on five video-understanding benchmarks for Qwen3.5 at 64 input frames. We bold the best result for each backbone and benchmark.
Method
Qwen3.5-0.8B
Qwen3.5-4B
Qwen3.5-9B
Qwen3.5-27B
16
32
64
16
32
64
16
32
64
16
32
64
Vanilla
37.8
40.1
41.6
48.0
49.8
53.0
49.0
51.3
54.4
54.3
56.8
59.5
SFT
38.6
40.9
41.8
48.5
50.7
53.6
49.8
52.2
55.2
54.8
56.1
59.8
GRPO
39.6
41.9
43.4
49.8
51.8
54.1
50.5
52.9
55.5
55.7
57.9
60.2
OPSD
39.1
41.5
43.5
49.1
51.0
53.9
50.3
52.6
55.5
55.7
58.1
60.2
OWLET
42.9
45.2
46.9
53.3
54.6
58.2
54.2
57.1
59.8
57.9
60.4
64.5
Table 2: Macro-average accuracy (%) over the five benchmarks for Qwen3.5, at 16 , 32 , and 64 input frames. We bold the best result for each backbone and frame budget.
Figure 4: Ablation results for Qwen3.5-9B across different frame budgets.
Figure 5: Training dynamics. Qwen3.5-9B with 64 student frames. (a) Hidden divergence, measured by response-end 1−CKA , for OPSD and OWLET. (b) Output KL between student and teacher views for OPSD and OWLET. (c) Rollout accuracy on the training set for OPSD and OWLET. (d) Cumulative absolute log-probability gap at the top 10% of response tokens for OPSD, GRPO, and OWLET; the dotted line marks the uniform baseline. Diagnostic samples are spaced 10 steps apart and accuracy samples 50 steps apart. Thin dashed lines connect sampled trajectories; thick lines show rolling means.
Appendix figures & tables47 assets
Supplementary material from the paper’s appendix.
Appendix
View
Uniform frames
Ranked frames
Window size
Total
16-frame budget
Vanilla
16
—
—
16
Key frames
—
16
—
16
Key windows
—
4
4
16
Global with local
8
2
4
16
32-frame budget
Appendix
Table 3: Composition of the four input views defined in Section 3.1 at total budgets of 16, 32, and 64 frames. Ranked frames are selected by question-conditioned relevance and used directly in Key frames or as centers of disjoint windows in Key windows and Global with local. All four views use the same total number of frames within each budget.
Figure 6: Teacher-view comparison on the Qwen3-VL model family. (a, b) Non-thinking variants. (c, d) Thinking variants. The left column reports accuracy on all questions; the right column reports accuracy on the global-temporal subset. Each panel compares the four input views defined in Table 3 on the 2B, 4B, and 8B models. Bars are labeled with accuracy in percent, and all four panels use the same vertical scale. In both modes and at all three model sizes, evidence-only views improve overall accuracy while reducing global-temporal accuracy relative to vanilla; global context with local windows improves both metrics.
Stage
Criterion
Excluded / reserved
Remaining
Source aggregation
four public sources
—
104,257
Field validation
non-empty question and answer
812
103,445
Option validation
unique labels, single correct option
1,134
102,311
Video validation
file exists and decodes end to end
1,313
100,998
Temporal validation
sufficient evidence candidates
4,293
96,705
Held-out diagnostics
evaluation split of Section 3
1,000
95,705
Appendix
Table 4: Construction pipeline of Video-Evidence-96K. Validation rows report excluded records and the number remaining after each check. The final row reserves 1,000 valid records for diagnostic evaluation, leaving 95,705 training records.
Source
Records
Share
LLaVA-Video-178K
70,468
73.63%
NExT-QA
7,542
7.88%
Perception
6,345
6.63%
STAR
11,350
11.86%
Total
95,705
100.00%
Appendix
Table 5: Composition of the Video-Evidence-96K training split.
Global-temporal
Local-evidence
Split
Total
Records
Percentage
Records
Percentage
Training
95,705
20,137
21.04%
75,568
78.96%
Held-out diagnostics
1,000
500
50%
500
50%
Appendix
Table 6: Composition of Video-Evidence-96K by question type, following Appendix D.3 . Counts are question–video records, and each percentage uses the total of its own split as the denominator.
Setting
Value
Optimization (shared by GRPO, OPSD, and OWLET )
Learning rate
1×10−6 , constant, no warm-up
Prompts per step / rollouts per prompt G
16 / 8
Mini-batch size
16
Training steps
1,000
Clipping range ϵ
0.2
Appendix
Table 7: Training and evaluation hyper-parameters.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3.5-0.8B
Vanilla
45.1
49.4
44.9
17.8
14.4
15.6
34.2
37.8
SFT
45.5
50.8
45.9
18.5
14.8
16.1
34.7
38.6
GRPO
46.8
50.7
47.0
19.6
16.3
17.5
35.8
39.6
OPSD
46.2
50.4
47.5
18.2
14.9
16.1
35.1
39.1
OWLET
50.3
56.4
50.6
20.5
17.6
18.6
38.9
42.9
Appendix
Table 8: Accuracy (%) for Qwen3.5 in non-thinking mode at 16 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3.5-0.8B
Vanilla
48.2
51.8
46.4
17.7
14.7
15.8
38.1
40.1
SFT
48.8
52.4
46.9
19.5
15.5
16.9
39.4
40.9
GRPO
50.1
54.8
48.2
19.5
16.4
17.5
38.9
41.9
OPSD
49.6
53.4
48.0
19.5
15.5
16.9
39.7
41.5
OWLET
53.9
57.8
53.0
22.2
18.0
19.4
42.0
45.2
Appendix
Table 9: Accuracy (%) for Qwen3.5 in non-thinking mode at 32 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3.5-0.8B
Vanilla
49.4
51.8
48.9
17.4
14.9
15.8
41.9
41.6
SFT
49.5
52.2
49.5
17.2
14.9
15.7
42.1
41.8
GRPO
50.8
53.9
50.4
19.2
18.1
18.5
43.3
43.4
OPSD
50.9
53.9
51.4
18.7
16.4
17.2
44.1
43.5
OWLET
54.9
58.6
55.9
22.6
18.4
19.9
45.2
46.9
Appendix
Table 10: Accuracy (%) for Qwen3.5 in non-thinking mode at 64 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Frames
Method
Short
Medium
Long
All
16
Vanilla
57.1
41.5
36.7
45.1
SFT
57.4
41.3
37.9
45.5
GRPO
59.5
42.7
38.2
46.8
OPSD
57.9
42.5
38.3
46.2
OWLET
61.8
47.9
41.3
50.3
32
Vanilla
58.7
46.7
39.2
48.2
Appendix
Table 11: Video-MME accuracy (%) by video duration (Qwen3.5-0.8B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Vanilla
69.1
54.6
48.1
57.2
SFT
70.3
55.9
48.6
58.3
GRPO
69.8
56.2
50.5
58.8
OPSD
70.3
56.1
50.6
59.0
OWLET
73.7
59.9
55.0
62.8
32
Vanilla
73.4
58.8
52.1
61.4
Appendix
Table 12: Video-MME accuracy (%) by video duration (Qwen3.5-4B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Vanilla
69.1
54.8
50.7
58.2
SFT
70.2
54.0
52.4
58.9
GRPO
70.4
58.0
52.1
60.2
OPSD
70.7
56.7
52.6
60.0
OWLET
71.2
60.8
54.6
62.2
32
Vanilla
75.1
58.8
52.1
62.0
Appendix
Table 13: Video-MME accuracy (%) by video duration (Qwen3.5-9B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Vanilla
74.5
63.5
56.5
64.8
SFT
74.3
62.8
57.1
64.7
GRPO
75.6
64.5
57.9
66.0
OPSD
75.4
64.1
58.0
65.8
OWLET
77.9
67.7
59.0
68.2
32
Vanilla
79.7
67.5
59.6
68.9
Appendix
Table 14: Video-MME accuracy (%) by video duration (Qwen3.5-27B). All averages the three splits. We bold the best result for each frame budget and duration.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Vanilla
45.2
53.1
42.9
14.4
14.3
14.3
34.8
38.1
SFT
45.8
53.1
43.4
15.4
14.9
15.1
35.8
38.6
GRPO
47.0
55.4
44.2
15.8
15.9
15.9
37.2
39.9
OPSD
46.5
53.2
45.3
15.9
15.5
15.6
36.7
39.5
OWLET
52.4
61.2
50.7
19.9
18.1
18.7
40.6
44.7
Appendix
Table 15: Accuracy (%) for Qwen3-VL in non-thinking mode at 16 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Vanilla
49.4
53.8
44.3
16.1
15.3
15.6
37.1
40.0
SFT
50.3
54.8
46.9
17.1
16.4
16.6
38.8
41.5
GRPO
51.8
55.0
45.4
18.2
16.6
17.2
40.2
41.9
OPSD
51.0
55.3
46.4
18.1
16.4
17.0
38.1
41.6
OWLET
56.2
60.9
54.9
22.7
18.1
19.7
45.3
47.4
Appendix
Table 16: Accuracy (%) for Qwen3-VL in non-thinking mode at 32 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Vanilla
49.2
53.9
45.1
16.8
17.3
17.1
40.3
41.1
SFT
50.1
54.6
46.8
18.1
18.0
18.0
41.8
42.3
GRPO
51.4
55.0
47.6
17.9
19.7
19.1
41.7
42.9
OPSD
50.7
54.0
47.1
18.1
18.2
18.2
42.1
42.4
OWLET
58.3
62.2
56.0
22.8
20.2
21.1
48.2
49.2
Appendix
Table 17: Accuracy (%) for Qwen3-VL in non-thinking mode at 64 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Frames
Method
Short
Medium
Long
All
16
Vanilla
58.4
43.1
34.2
45.2
SFT
59.0
43.4
34.9
45.8
GRPO
59.8
44.9
36.3
47.0
OPSD
59.9
44.6
35.1
46.5
OWLET
64.4
51.4
41.4
52.4
32
Vanilla
65.1
48.0
35.0
49.4
Appendix
Table 18: Video-MME accuracy (%) by video duration for Qwen3-VL in non-thinking mode (Qwen3-VL-2B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Vanilla
62.2
45.9
37.9
48.7
SFT
64.2
47.7
38.4
50.1
GRPO
64.7
46.8
39.8
50.4
OPSD
64.4
47.0
39.2
50.2
OWLET
69.1
58.4
49.2
58.9
32
Vanilla
68.9
50.4
43.8
54.3
Appendix
Table 19: Video-MME accuracy (%) by video duration for Qwen3-VL in non-thinking mode (Qwen3-VL-4B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Vanilla
68.4
55.1
50.0
57.9
SFT
69.6
57.2
50.9
59.2
GRPO
70.9
54.9
51.9
59.2
OPSD
69.5
56.4
51.4
59.1
OWLET
72.2
57.8
52.2
60.7
32
Vanilla
74.0
58.0
54.1
62.1
Appendix
Table 20: Video-MME accuracy (%) by video duration for Qwen3-VL in non-thinking mode (Qwen3-VL-8B). All averages the three splits. We bold the best result for each frame budget and duration.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Base model
49.3
60.5
45.1
17.8
15.3
16.2
35.1
41.3
SFT
51.0
61.5
46.0
19.4
16.8
17.7
37.1
42.7
GRPO
50.9
61.8
46.0
18.3
15.2
16.3
36.3
42.3
OPSD
50.0
61.3
46.6
19.9
16.6
17.8
35.2
42.2
OWLET
54.2
65.9
52.8
22.1
17.4
19.1
43.0
47.0
Appendix
Table 21: Accuracy (%) for Qwen3-VL in thinking mode at 16 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Base model
52.2
61.0
51.0
20.1
16.1
17.5
38.0
44.0
SFT
53.3
62.1
52.3
20.3
16.7
18.0
38.1
44.8
GRPO
53.5
61.2
51.9
21.9
17.4
19.0
39.1
44.9
OPSD
53.1
62.1
52.9
21.3
17.4
18.8
39.1
45.2
OWLET
57.4
67.2
55.0
25.5
19.4
21.5
46.0
49.4
Appendix
Table 22: Accuracy (%) for Qwen3-VL in thinking mode at 32 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Backbone Model
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
Logic
Rel.
All
Qwen3-VL-2B
Base model
55.5
63.1
51.4
20.1
18.1
18.8
46.0
47.0
SFT
56.4
65.2
52.1
20.6
19.5
19.9
46.2
48.0
GRPO
56.7
64.0
52.4
21.2
19.2
19.9
47.3
48.1
OPSD
56.5
64.7
52.0
20.9
18.8
19.5
47.1
48.0
OWLET
60.3
66.8
55.6
25.6
21.3
22.8
49.6
51.0
Appendix
Table 23: Accuracy (%) for Qwen3-VL in thinking mode at 64 input frames, with the two Video-MME 2 question groups reported separately. Logic and Rel. are the group types and All combines them by question count. We bold the best result for each backbone and benchmark.
Frames
Method
Short
Medium
Long
All
16
Base model
61.5
46.3
40.1
49.3
SFT
63.5
48.5
41.1
51.0
GRPO
63.0
47.5
42.2
50.9
OPSD
62.6
46.9
40.5
50.0
OWLET
66.0
52.8
43.8
54.2
32
Base model
65.2
50.4
41.0
52.2
Appendix
Table 24: Video-MME accuracy (%) by video duration for Qwen3-VL in thinking mode (Qwen3-VL-2B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Base model
65.4
52.1
46.0
54.5
SFT
67.1
53.4
47.5
56.0
GRPO
66.9
52.9
46.7
55.5
OPSD
66.2
53.0
47.4
55.5
OWLET
70.3
56.5
50.6
59.1
32
Base model
72.4
56.1
48.2
58.9
Appendix
Table 25: Video-MME accuracy (%) by video duration for Qwen3-VL in thinking mode (Qwen3-VL-4B). All averages the three splits. We bold the best result for each frame budget and duration.
Frames
Method
Short
Medium
Long
All
16
Base model
66.1
52.6
47.2
55.3
SFT
67.0
52.8
47.7
55.8
GRPO
68.4
54.4
48.2
57.0
OPSD
67.8
53.5
47.8
56.4
OWLET
71.4
58.8
51.2
60.5
32
Base model
73.2
57.2
53.2
61.2
Appendix
Table 26: Video-MME accuracy (%) by video duration for Qwen3-VL in thinking mode (Qwen3-VL-8B). All averages the three splits. We bold the best result for each frame budget and duration.
Figure 7: Additional motivation diagnostics across four configurations. Each subfigure contains the same two diagnostics: the left panel reports the cumulative absolute teacher–student log-probability gap, with the marked point giving the response-token fraction that carries 50% of the gap; the right panel reports output KL and hidden divergence during 300 steps of output-only distillation.
Figure 8: Ablation results for Qwen3.5-0.8B. Panels (a)–(c) use 16 , 32 , and 64 student frames, respectively. The five axes report accuracy (%) on Video-MME, MVBench, LongVideoBench, Video-MME 2, and MLVU. The legend defines the key-window, w/o ECR, and w/o SSA variants described in the ablation text; evaluation uses uniformly sampled student frames.
Figure 9: Ablation results for Qwen3.5-4B. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 10: Ablation results for Qwen3.5-27B. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 11: Ablation results for Qwen3-VL-2B in non-thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 12: Ablation results for Qwen3-VL-2B in thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 13: Ablation results for Qwen3-VL-4B in non-thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 14: Ablation results for Qwen3-VL-4B in thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 15: Ablation results for Qwen3-VL-8B in non-thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Figure 16: Ablation results for Qwen3-VL-8B in thinking mode. Panels (a)–(c) correspond to 16 , 32 , and 64 uniformly sampled student frames.
Method
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
OPSD
66.1
68.7
62.9
26.7
53.3
55.5
OPSD-JSD
65.3
68.2
61.5
25.9
53.7
54.9
OPSD+GRPO
66.3
67.8
63.3
27.0
52.7
55.4
OPSD-Anneal
67.1
68.5
63.4
27.1
55.3
56.3
OWLET
70.7
72.8
65.2
28.9
61.6
59.8
Appendix
Table 27: Accuracy (%) of on-policy distillation variants and OWLET on Qwen3.5-9B at 64 student frames. All methods use the same teacher view, training data, and student rollout budget. Avg. is the unweighted mean over the five benchmarks. The best result in each column is shown in bold.
Teacher θˉ
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
EMA (decay 0.99 )
69.3
70.5
64.7
27.9
60.8
58.6
Fixed (default)
70.7
72.8
65.2
28.9
61.6
59.8
Appendix
Table 28: Effect of the teacher parameterization on Qwen3.5-9B at 64 input frames. The fixed teacher retains the initial checkpoint throughout training, while the EMA teacher averages student parameters with decay 0.99 under the same teacher view and training configuration. Scores are accuracies (%), and Avg. is the unweighted mean over the five benchmarks.
Advantage
Steps 0–250
250–500
500–750
750–1,000
Ai>0
28.5
21.8
10.6
4.7
Ai<0
7.2
5.8
3.8
2.2
Ai=0
64.3
72.4
85.6
93.1
Appendix
Table 29: Fractions (%) of rollouts with positive, negative, and zero advantages across training phases on Qwen3.5-9B at 64 input frames.
Category
Top decile
Bottom decile
Examples
Temporal markers
17.3
6.2
after, then, at the end
Event verbs and nouns
20.7
5.2
pours, jumps, collision
Objects, places, attributes
21.0
7.5
kitchen, red container, three
Answer-committing tokens
15.3
5.9
option letters, final numbers
Function words, connectives, and auxiliaries
4.8
23.1
the, is, and, therefore
Punctuation and formatting
3.1
25.4
commas, line breaks
Appendix
Table 30: Composition (%) of tokens in the top and bottom weight deciles among positively advantaged responses. Each response token is assigned to one of seven mutually exclusive categories; tokens that do not match the predefined rules are assigned to Others .
λSSA
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
0 (w/o SSA)
69.6
71.1
64.6
28.4
59.9
58.7
0.01
70.5
71.1
64.8
27.9
59.5
58.8
0.02
70.3
71.5
65.3
28.9
59.9
59.2
0.05 (default)
70.7
72.8
65.2
28.9
61.6
59.8
0.1
70.9
71.3
64.8
28.3
61.9
59.4
Appendix
Table 31: Sensitivity to the SSA weight on Qwen3.5-9B in non-thinking mode with 64 uniformly sampled student frames. Scores are accuracies (%), and Avg. is the unweighted mean over the five benchmarks. All other training settings are held constant; λSSA=0 disables SSA while retaining ECR, and 0.05 is the default weight.
Figure 17: Global-temporal case. Responses from Vanilla and checkpoints trained with SFT, GRPO, OPSD, and OWLET to the same video input and question. Vanilla and GRPO select D, and SFT and OPSD select B; OWLET recovers the correct camping–queuing–cheering order and selects C.
Figure 18: Local-evidence case. Responses to the same video input and question about the object visible at a specified cooking step. Vanilla, SFT, and GRPO select C (green onion), and OPSD selects A (knife); OWLET identifies the pot and selects the correct answer B.
Method
Rollout (s/step)
Update (s/step)
Total (s/step)
Total (h)
GRPO
406.99
507.30
914.29
254.0
OPSD
410.89
437.11
848.00
235.6
OWLET
434.31
514.24
948.55
263.5
Appendix
Table 32: On-policy training cost on eight H200 GPUs for Qwen3.5-9B at 64 student frames, averaged over the first 1,000 updates. Update covers the non-rollout portion of each timed step. Total (s/step) is the sum of Rollout and Update; Total (h) gives the accumulated step time.
Figure 19: Response length during the first 1,000 training steps of OWLET with 64 student frames: (a) Qwen3.5-9B in non-thinking mode, (b) Qwen3-VL-8B-Instruct in non-thinking mode, and (c) Qwen3-VL-8B-Thinking. Thin, light curves show the mean number of non-padding generated tokens across all 128 responses at each step; thick curves show 25-step trailing moving averages, using the available history for the first 24 steps. Thinking lengths include the generated reasoning trace and final answer. The panels use separate vertical scales, and all per-step means are shown.
Model
Mode
Steps 1–100
Steps 901–1,000
Overall
Change (%)
Qwen3.5-9B
Non-thinking
42.4
33.9
40.4
−20.0
Qwen3-VL-8B
Non-thinking
77.5
72.0
75.2
−7.1
Qwen3-VL-8B
Thinking
145.6
123.7
135.2
−15.0
Appendix
Table 33: Mean response length in generated tokens for the three runs in Figure 19 . Overall averages all 1,000 steps. Change is the percentage difference between the final and first 100-step means, computed before rounding.
Method
Overall
Global-temporal
Local-evidence
Vanilla
50.5
51.8
49.2
Key Windows
57.8
50.2
65.4
GRPO
55.8
53.4
58.2
OPSD
54.5
52.4
56.6
OWLET
59.2
58.6
59.8
Appendix
Table 34: Accuracy (%) on the held-out diagnostic split before and after post-training, overall and by question type, for Qwen3.5-9B at 64 input frames. The Vanilla and Key Windows rows use the same initial checkpoint with uniformly sampled and question-conditioned Key Windows views, respectively; GRPO, OPSD, and OWLET are evaluated at their final checkpoints under the uniformly sampled student view.
Training steps
Video-MME
MVBench
LongVideoBench
Video-MME 2
MLVU
Avg.
1,000 (main text)
70.7
72.8
65.2
28.9
61.6
59.8
1,200
70.5
73.1
65.5
28.4
60.8
59.7
1,500
71.0
73.6
65.5
28.6
60.5
59.8
1,800
72.3
73.0
64.8
29.5
61.0
60.1
2,000
72.0
73.6
66.1
29.2
62.7
60.7
2,200
70.9
73.3
66.9
29.1
62.9
60.6
Appendix
Table 35: Per-benchmark accuracy (%) of OWLET checkpoints under an extended training schedule, for Qwen3.5-9B at 64 input frames. Checkpoints are saved at nine steps between the main-text budget of 1,000 and 3,000 , spaced 200 – 300 steps apart, and each is evaluated with the protocol of Table 1 on the same five benchmarks; the 1,000 -step row reproduces the corresponding OWLET entry there. Avg. is the macro-average over the five benchmarks, computed before rounding. The last row gives the change from 1,000 to 3,000 steps.
Video reasoning has advanced with large multimodal models (LMMs), yet their inference is often a single pass that returns an answer without verifying whether the reasoning is evidence-aligned. We introduce Reinforce to Learn, Elect to Reason (RLER), a dual paradigm that decouples learning to produce evidence from obtaining a reliable answer. In RLER-Training, we optimize the policy with group-relative reinforcement learning (RL) and 3 novel task-driven rewards: Frame-sensitive reward grounds reasoning on explicit key frames, Think-transparency reward shapes readable and parsable reasoning traces, and Anti-repetition reward boosts information density. These signals teach the model to emit structured, machine-checkable evidence and potentiate reasoning capabilities. In RLER-Inference, we apply a train-free orchestrator that generates a small set of diverse candidates, parses their answers and cited frames, scores them by evidence consistency, confidence, transparency, and non-redundancy, and then performs a robust evidence-weighted election. This closes the loop between producing and using evidence, improving reliability and interpretability without enlarging the model. We comprehensively evaluate RLER against various open-source and RL-based LMMs on 8 representative benchmarks. RLER achieves state of the art across all benchmarks and delivers an average improvement of 6.3% over base models, while using on average 3.1 candidates per question, indicating a favorable balance between compute and quality. The results support a simple thesis: making evidence explicit during learning and electing by evidence during inference is a robust path to trustworthy video reasoning.
Songyuan Yang, Weijiang Yu, Jilin Ma +5
National University of Defense Technology · Sun Yat-sen University
On-policy self-distillation (OPSD) has recently emerged as an effective post-training paradigm that improves policy optimization through dense token-level supervision from a privileged self-teacher. Despite its promise, OPSD remains largely underexplored for Video Large Language Models (Video-LLMs). Existing methods typically construct privileged teachers by augmenting their context with additional information while keeping the primary input unchanged for both teacher and student. Video reasoning, however, offers a distinct source of privileged supervision within the primary input itself: long videos contain substantial temporal redundancy, and only a small subset of frames provides the evidence necessary to answer a question. Building on this observation, we present Video-OPSD, an OPSD framework that exploits privileged visual evidence for both self-teacher construction and knowledge transfer. First, our Evidence-Grounded Self-Teacher conditions the teacher exclusively on annotated evidence frames while the student continues to reason over the complete video. This focused visual input enables the teacher to provide more informative supervision. Second, our Evidence-Guided Token Optimization adaptively weights token-level distillation according to each reasoning token's reliance on privileged visual evidence, thereby emphasizing perceptually grounded reasoning. Experiments across video understanding and reasoning benchmarks show that Video-OPSD consistently improves upon Standard OPSD across multiple backbones and achieves performance comparable to GRPO while requiring substantially less training time, establishing an effective and efficient post-training approach for Video-LLMs.
Ziyue Wang, Shiqi Huang, Weiwen Xu +2
School of Electrical and Electronic Engineering, Nanyang Technological University · The Chinese University of Hong Kong
Streaming video understanding requires Video Large Language Models (Video-LLMs) to reason over continuous visual streams under causal constraints. As the visual history grows, a bounded visual?processing budget requires evidence selection that balances temporal recency with query relevance. Recent-only selection excludes potentially relevant historical evidence, whereas Semantic-only retrieval can displace useful recent context when relevance scores are ambiguous. We introduce WRWS (When to Retrieve, When to Stay), a training-free framework for uncertainty-adaptive evidence allocation. A lightweight external vision-language encoder scores query relevance across the observed history, while an adaptive allocation module uses the normalized entropy of the similarity distribution as a proxy for retrieval uncertainty. WRWS favors semantic retrieval when relevance cues are reliable and strengthens the recency prior under uncertainty. Following a retrieve-first, encode-later pipeline, WRWS selects evidence before target-model visual encoding, such that only the selected observations are processed by the costly target Video-LLM. Experiments across four Video-LLM families and multiple model scales demonstrate competitive accuracy on StreamingBench and OVO-Bench. In our efficiency evaluation, WRWS reduces average vision-to-answer time to 47.93% of the state-of-the-art method. Code will be released.