Under a fixed frame budget, video understanding with vision-language models (VLMs) depends on whether the sampled frames provide sufficient evidence for the question. Sufficient frame evidence supports direct answering, whereas less informative observations may benefit from reasoning that connects the available cues. We introduce EVAThink, which enhances video understanding through evidence sufficiency-guided training for adaptive thinking. First, Recoverable Evidence Gap Mining constructs evidence-dense and evidence-sparse video views that share the same question and answer under an identical frame budget. It retains pairs where sparse views yield lower direct-answer accuracy than dense views but benefit from explicit reasoning, providing supervision for response-mode selection. Second, Evidence-Guided Two-Stage Reinforcement Learning uses stage-specific rewards to guide training with evidence-based mode preferences and then refine response efficiency, while first-token advantage regularization mitigates mode collapse. Trained on only 6.7K curated samples, EVAThink consistently improves upon the base model across video understanding benchmarks under both 32- and 64-frame budgets, while outperforming strong baselines on average. The code, model, and datasets are available at EVAThink.
Figures & tables
Figure 1: Motivation for evidence sufficiency-guided training for adaptive thinking. For the same question and frame budget, sufficient frame evidence supports direct answering, while thinking can help make better use of limited frame evidence.
Figure 2: Overview of EVAThink. Recoverable Evidence Gap Mining constructs the evidence-contrastive pairs through video cropping or concatenation, retaining pairs with a direct-answer accuracy gap and a reasoning gain on the sparse view. Evidence-Guided Two-Stage Reinforcement Learning first trains response-mode selection using preferences derived from evidence-contrastive pairs, then refines response efficiency on the remaining pairs, using stage-specific rewards and first-token advantage regularization.
Figure 3: Data pool.
Model
Train. Size
Video-MME
MVBench
LongVideoBench
MMVU(mc)
Video-MMMU
VSI-Bench
Avg.
Results reported under the respective evaluation protocols (different visual budgets and evaluation settings)
Temporal-RLT
32K
57.6
68.1
-
65.0
-
-
-
Video-RTS
6K
63.0
-
56.6
66.4
52.7
-
-
VITAL
182K
64.1
-
-
68.7
54.2
41.8
-
LongVILA-R1
206K
65.1
67.6
58.0
-
-
-
-
LOVE-R1
192.5K
66.2
66.6
60.1
-
-
-
-
Table 1: Performance across video understanding benchmarks. Bold and underlined values indicate the best and second-best scores within each fixed-budget group, respectively. Avg. is the unweighted mean over all six benchmarks and is reported only for complete results. † AdaThinkV allows up to 256 input frames.
Model
Think Mode
Avg.
Tok.
EVAThink
Direct
63.3
8
EVAThink
Think Only
62.0
560
EVAThink
Auto Think
65.6
213
Table 2: Effect of adaptive thinking under the 64-frame setting. Tok. denotes the average number of generated tokens per response.
Model
Think Ratio
Overall Acc. (%)
Dense
Sparse
EVAThink
22%
83%
65.3
Table 3: Evaluation on held-out dense/sparse pairs sharing the same question and answer, evaluated under the 64-frame setting.
Model
MMBench-En
MMBench-Cn
MME-P
MME-C
AI2D
POPE
Rel.
Qwen3-VL-8B-Instruct
83.7
82.2
1671.9
546.4
81.9
88.2
100%
EVAThink
85.1
82.9
1723.1
562.5
82.5
88.8
101.7%
Table 4: More results across image understanding tasks. “Rel.” denotes the average relative performance across benchmarks, computed as the mean percentage normalized by the base model (Qwen3-VL-8B-Instruct) set to 100%.
Method
Video-MME
MVBench
LongVideoBench
MMVU(mc)
Video-MMMU
VSI-Bench
Avg.
Tok.
w/o Recoverability filtering
65.9
65.6
60.2
72.3
67.2
58.5
65.0
247
w/o Evidence-gap & Recoverability filtering
64.8
65.1
59.5
69.8
63.3
58.2
63.5
198
w/o Stage 1 mode preference reward
63.2
62.7
56.4
68.2
59.7
56.7
61.2
392
w/o Stage 2 direct-answer bonus reward
66.0
65.9
60.5
72.6
68.1
58.6
65.3
465
w/o First-token advantage regularization
66.1
65.2
60.2
72.2
63.8
60.1
64.6
52
w/o Stage1 training
66.2
66.8
60.0
71.5
63.2
58.9
64.4
80
Table 5: Ablation study under the 64-frame setting. RL optimization settings are held fixed; data-curation changes are described in the text.
Figure 4: Effect of first-token advantage regularization on policy entropy during Stage 1 (left) and Stage 2 (right). Without regularization, entropy rapidly collapses in both stages, accompanied by a shift to direct answering for all rollouts (response-mode collapse).
Figure 5: Case studies of adaptive thinking: EVAThink avoids unnecessary reasoning (left), prevents reasoning-induced errors through direct answering (middle), and resolves errors through evidence-grounded reasoning (right).
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Data source
Selection criterion
# Selected samples
LLaVA-Video-178K
Academic subset
28.9K
Molmo2-VideoCapQA
Random subset
21.9K
VideoMarathon
Random subset
4.5K
LongVILA
All
29.1K
SR-91K
Multiple-choice subset
18.4K
Holmes-train
All
1.6K
Appendix
Table 6: Composition of the data pool.
USER PROMPT FOR DIRECT ANSWER
You are given a video. Please answer the question directly. Question: [Question] [Options] Your final answer must be enclosed in <answer></answer> . - For a multiple-choice question, put only the option letter inside <answer></answer> . Examples: <answer> A </answer>
Appendix
Table 7: User prompt for direct answer in the data curation.
USER PROMPT FOR THINKING
You are given a video.
Do not answer immediately. Think step by step before answering.
If the frames are temporally jumpy, the evidence is scattered, or the visible content is not sufficient for an immediate answer, carefully reason over the observed cross-frame clues and integrate the evidence before giving your answer.
Do not invent details that are not supported by the visible frames.
Note:
Work through your reasoning first, then give your final answer.
Appendix
Table 8: User prompt for thinking in the data curation.
USER PROMPT
You are a video evidence localizer.
You will be given a video constructed from exactly 32 frames uniformly sampled from the original video and then arranged in chronological order.
The frame ids are defined by their order in this 32-frame video:
- The first sampled frame is Frame 0.
- The last sampled frame is Frame 31.
Your task is not to answer the question.
Appendix
Table 9: Prompt for video evidence segment annotation, which instructs the model to identify the minimal sampled frame range supporting the ground-truth answer.
Data source
Stage 1 (pairs)
Stage 2 (pairs)
LLaVA-Video-178K
317
628
Molmo2-VideoCapQA
325
529
VideoMarathon
51
169
LongVILA
329
720
SR-91K
514
1682
Holmes-train
28
67
Appendix
Table 10: Composition of the evidence-contrastive training pairs for the two stages.
USER PROMPT
Answer the question based on the video. First decide whether explicit written reasoning is necessary. Choose exactly one of the following two response modes. Your response must begin with either <answer> or Alright with no other opening.
1. Direct-answer mode
Use this mode when the visual evidence clearly supports the final answer and no written reasoning is needed. Begin directly with <answer> and output only the final answer block.
Example: <answer> A </answer>
2. Evidence-grounded reasoning mode
Use this mode only when the answer is not immediately clear, or when the relevant evidence is temporally sparse, discontinuous, or scattered across multiple frames. Begin with Alright , briefly connect the relevant observed clues, and then provide the final answer block. Include only as much reasoning as necessary to determine the answer.
Appendix
Table 11: Training prompt for response mode selection, which instructs the model to adaptively select between direct and thinking modes.
Model
Think Mode
Video-MME
MVBench
LongVideoBench
MMVU(mc)
Video-MMMU
VSI-Bench
Avg.
Tok.
EVAThink
Direct
66.2
66.4
59.8
70.6
56.8
59.9
63.3
8
EVAThink
Think Only
60.8
64.2
57.4
71.2
67.1
51.4
62.0
560
EVAThink
Auto Think
66.3
66.5
60.7
72.8
67.4
59.9
65.6
213
(think ratio)
(34.4%)
(5.1%)
(46.1%)
(59.4%)
(85.7%)
(0.0%)
Appendix
Table 12: Effect of adaptive thinking under the 64-frame setting. Tok. denotes the average number of generated tokens per response.
Hyperparameter
Value
Batch size
32
Rollouts per prompt
8
Policy updates per step
1
Optimizer
AdamW
Learning rate
1e-6 (constant)
Gradient clipping
1.0
Appendix
Table 13: Training Hyperparameters.
Figure 6: Case studies of EVAThink: think mode (1).
Figure 7: Case studies of EVAThink: think mode (2).
Figure 8: Case studies of EVAThink: direct mode (1).
Figure 9: Case studies of EVAThink: direct mode (2).