PaperlayerView exact revision on arXiv

Source-linked AI summary

Efficient Reasoning Distillation: Small Video-Language Models via Synthetic CoT and Difficulty-Aware Fine-Tuning

Mantek Singh, Jeshwanth Challagundla, Siddharth Raina, Jasmin Jarsania

arXiv:2609.16255v1cs.LGcs.AIcs.CV

TL;DR

Compact VLMs need stronger video-question reasoning while remaining suitable for resource-constrained deployment. The paper fine-tunes a 2B model with uncertainty-selected examples and synthetic CoT, reporting generalization across benchmarks and performance rivaling models up to four times larger.

  • Problem

    Current VLMs can be prohibitively large for resource-constrained deployment, motivating compact models with strong VideoQA reasoning.

    Source: §1 · p. 1

  • Method

    The method fine-tunes a 2B student on about 900 uncertainty-selected examples augmented with synthetic CoT rationales from a 4B teacher, using Answer-then-Rationale formatting.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    The fine-tuned model generalizes across CinePile, ActivityNet-QA, and MLVU while achieving performance that rivals models up to 4× larger.

    Source: §6 · p. 11

  • Takeaways & Limitations

    The results support a practical route to reasoning-rich compact VLMs using targeted synthetic data and fine-tuning.

    Source: §5 · pp. 9–10

    Further validation across architectures is needed, and GPT-generated options in the ActivityNet-QA evaluation could influence outcomes.

    Source: Limitations · p. 10

Abstract

from arXiv · show

We present an efficient method to distill reasoning capabilities into compact video-language models (VLMs) for video question answering (VideoQA). Our approach fine-tunes a 2B-parameter model using only $\sim$900 uncertainty-selected examples, each augmented with synthetic chain-of-thought (CoT) rationales generated by a 4B teacher. Despite its minimal compute cost - under two hours on a single A100 GPU - our method enables the 2B model to outperform VLMs up to 4$\times$ larger, and generalize across CinePile, ActivityNet-QA, and MLVU, approaching the performance of its own 4B teacher. A key finding is that placing CoT rationales after the answer - contrary to standard prompting - substantially improves reasoning in compact models. This insight challenges prevailing CoT conventions and reveals new alignment strategies under limited model capacity. Our findings offer a practical blueprint for training deployable, reasoning-rich VLMs suited for mobile and edge applications.

1 Introduction

The paper targets compact VLMs that can reason about video questions, using synthetic rationales and uncertainty-guided selection to make fine-tuning sample-efficient. It reports gains from synthetic CoT, top-2 margin sampling, and placing rationales after answers, with generalization beyond CinePile.

1 Introduction

  • The 2B InternVL2-2B student distills reasoning from a 4B InternVL2-4B teacher using synthetic CoT rationales and hard-sample selection.The selection targets examples with high model-estimated difficulty or a small top-2 prediction margin.Source: §1 · p. 2
  • Top-2 margin sampling consistently outperforms the other selection strategies, while placing rationales after answers substantially boosts reasoning in small models.The rationale order is contrary to standard CoT prompting.Source: §1 · p. 2
  • Fine-tuning on only 900 examples (<0.5% of training data) yields large performance gains, with improvements generalizing across unseen datasets and domains.The introduction identifies CinePile as the primary benchmark and ActivityNet-QA and MLVU as generalization datasets.Source: §1 · p. 2
  • The authors present the combination of small-scale CoT distillation, uncertainty-guided sampling, and format-aware rationale design as a practical recipe for compact VLMs.They identify deployment under real-world constraints as the intended setting.Source: §1 · p. 2

2 Related Work

Related work frames this approach against synthetic-data and distillation methods, CoT reasoning and hard-sample selection, and existing VideoQA benchmarks. The paper focuses on concise reasoning supervision for compact models and evaluation spanning distinct video and task settings.

2.1 Synthetic Data and Knowledge Distillation

  • Small models under 3B can struggle with complex supervision, while prior video-reasoning distillation methods such as AoTD rely on extensive supervision.The paper contrasts this with its use of a 4B teacher and 900 training examples for practical edge deployment.Source: §2.1 · p. 2

2.2 Chain-of-Thought Reasoning and Hard Sample Selection

  • This work studies concise CoTs from a 4B teacher for a 2B model, including rationale positioning and format for long-form video QA.For selection, it uses margin-based uncertainty sampling and a hard-first approach, unlike typical easy-to-hard curriculum learning.Source: §2.2 · p. 3

2.3 VideoQA Datasets

  • The evaluation spans long-form, temporally rich CinePile, diverse web-video ActivityNet-QA, and compositional, multi-task MLVU.The paper distinguishes these settings from benchmarks focused on shorter coverage or other long-context tasks.Source: §2.3 · p. 3, Table 1 · p. 3

3 Methodology

The methodology selects uncertain examples, augments them with concise CoT rationales from a 4B teacher, and fine-tunes a 2B student. It evaluates CinePile performance and zero-shot generalization, using a resource-conscious frame setup.

3 Methodology

  • The pipeline combines hard-sample selection, synthetic CoT rationale generation, and efficient supervision to improve InternVL2-2B video reasoning.The methodology overview is shown in Fig. 1.Source: §3 · p. 3, Figure 1 · p. 4

3.1 Hard Sample Selection and Rationale Generation

  • Top-2 margin selection chooses examples with the smallest gap between the two highest predictions; the study uses 900 samples after performance saturated near 700.The 900 examples are less than 0.5% of the full dataset; rationales from the 4B teacher are capped at 20 tokens, with 85% validity in manual inspection.Source: §3.1 · pp. 3–4

3.2 Fine-Tuning and Evaluation

  • The adopted Answer-then-Rationale format places the answer before a concise rationale, after alternatives including rationale-first and rationale-as-input produced lower gains.The model is fine-tuned for three epochs with AdamW at a 2e-5 learning rate; evaluation includes CinePile and zero-shot ActivityNet-QA adapted with GPT-4o-generated distractors.Source: §3.2 · p. 4

3.3 Implementation Details

  • Fine-tuning uses 3 video frames, compared with 12 frames during inference, and completes in under two hours on a 4xA100 node.The implementation uses PyTorch and Hugging Face Transformers.Source: §3.3 · pp. 4–5

4 Experiments and Results

The experiments test whether uncertainty-based sample selection and synthetic CoT improve the efficiency and reasoning performance of a 2B video-language model. The strongest CinePile result approaches the 4B model, with additional gains reported on MLVU and zero-shot ActivityNet-QA.

4.3 Main Results on CinePile

  • 38.34% average accuracy lets InternVL2-2B (Top-2 Gap+CoT) exceed mPLUG-Owl3-8B (38.27%) and approach InternVL2-4B (39.89%), while remaining below LongVA-7B (41.04%).The reported CinePile gap to InternVL2-4B is 1.55 points.Source: Table 2 · p. 6, §4.3 · p. 8

4.2 Sample Efficiency Analysis

  • 34.85% Acc from 300 Top-2 Gap samples surpasses 31.02% Acc from 900 Random samples; targeted-strategy gains plateau around 700 samples.The sample-count curves compare Random, Top-2 Gap, and Top-2 Gap+CoT selection.Source: Figure 2 · p. 5, §4.2 · p. 7
  • 37.81% Acc with 600 Top-2 Gap+CoT samples is comparable to 36.34% Acc with 900 samples without CoT.The Top-2 Gap+CoT curve remains above the Top-2 Gap curve, and gains begin to saturate around 700 samples.Source: Figure 2 · p. 5, §4.2 · p. 7
  • Compared with InternVL2-8B, InternVL2-26B generated significantly longer CoTs with more reasoning-heavy tokens, potentially overwhelming smaller students and supporting the capacity-mismatch hypothesis about verbose CoTs hindering learning in low-capacity models.Source: §4.2 · p. 7
  • Unrelated, similar-length rationales degrade performance below baseline, supporting reasoning content—not added sequence length—as the source of CoT gains.The ablation replaces CoT with randomly generated unrelated text of similar length.Source: §4.2 · pp. 7–8, Table 4 · p. 8

4.3 Main Results on CinePile

  • 36.34% for Top-2 Gap exceeds Random (31.02%, +5.32 points) and Difficulty (33.68%, +2.66 points) on CinePile.All fine-tuning approaches outperform the 30.34% zero-shot baseline.Source: §4.3 · p. 8

4.4 Generalization to MLVU

  • 50.5% accuracy on a held-out MLVU subset is 5.5 points above the 2B base model (45.0%) and near the 4B teacher (51.0%).The reported margin to the teacher is 0.5 points.Source: Table 5 · p. 9, §4.4 · p. 9

4.5 Generalization to ActivityNet-QA (Zero-Shot Transfer)

  • 43.19% accuracy on synthetically adapted ActivityNet-QA exceeds the baseline by 2.99 points and fine-tuning without CoT by 1.73 points.Evaluation is zero-shot on a multiple-choice adaptation of the open-ended benchmark.Source: §4.5 · p. 9

5 Discussion and Limitations

The discussion attributes compact-model gains to the combination of targeted selection, synthetic CoT, and capacity-aware rationale design. It also limits the evidence by teacher reliability, confidence calibration, evaluation setup, and architecture coverage.

5 Discussion and Limitations

  • Fine-tuning InternVL2-2B on targeted CinePile data with synthetic CoT approached a 2× larger model and surpassed other large baselines.The authors describe this as an efficient path to model improvement.Source: §5 · pp. 9–10
  • Top-2 Gap selection improves data efficiency, while synthetic rationales amplify learning on high-uncertainty samples where explicit reasoning is most valuable.The authors connect the rationales to helping compact models disentangle temporal dependencies.Source: §5 · p. 10
  • For this task, a 4B teacher outperforms a 26B teacher, supporting the view that verbose, complex rationales can overwhelm a small student.The authors propose moderate-capacity teachers as a better quality–distillability trade-off.Source: §5 · p. 10
  • Answer-then-Rationale outperforms Rationale-then-Answer for compact models, motivating format design tailored to their capacity.The authors hypothesize that pre-answer reasoning burdens low-capacity models and that answer-first supervision supports optimization.Source: §5 · p. 10

Limitations

  • Synthetic CoT effectiveness is bounded by teacher quality, and confidence-based selection may be affected by miscalibration; ActivityNet-QA also uses GPT-generated options.The authors call for rationale-faithfulness auditing, particularly in high-stakes settings, and note that the synthetic evaluation layer could influence outcomes.Source: Limitations · p. 10
  • Evidence is limited to InternVL2-2B, with validation across architectures still needed and teachers above 4B infeasible for on-device generation.Extending the method to new architectures remains future work.Source: Limitations · pp. 10–11

6 Conclusion

The conclusion presents targeted synthetic CoT fine-tuning as a way to improve compact VLM video reasoning with limited data. It reports gains on CinePile, generalization to ActivityNet-QA and MLVU, and performance rivaling models up to four times larger.

6 Conclusion

  • InternVL2-2B is fine-tuned on about 900 high-uncertainty CinePile samples using concise CoT rationales generated by a 4B teacher.The conclusion describes the pipeline as using minimal data and compute.Source: §6 · p. 11
  • The method boosts InternVL2-2B on CinePile, generalizes to ActivityNet-QA, and reaches near-parity with its 4B teacher on MLVU.The conclusion also reports outperforming larger models on CinePile.Source: §6 · p. 11
  • A 4B teacher can outperform a 26B teacher in this setting, and the resulting method is presented as a route toward edge deployment of reasoning-rich VLMs.The authors report performance rivaling models four times larger with minimal data and compute.Source: §6 · p. 11

Related papers

Loading 2609.16255v1…