PaperlayerView exact revision on arXiv

Source-linked AI summary

Video-FLAIR: Not Whether to Reason, But How

Yogesh Kulkarni, Pooyan Fazli

arXiv:2608.26495v1cs.CV

TL;DR

Multimodal queries require different reasoning types, but many methods apply a uniform strategy, wasting computation on simple tasks and underserving complex ones. Video-FLAIR uses reinforcement learning and within-query mode comparisons to select among adaptive reasoning modes, improving benchmark accuracy while reducing token usage.

  • Problem

    Existing methods often apply one reasoning strategy across queries, causing unnecessary computation on simple tasks and insufficient reasoning on complex ones.

  • Method

    Video-FLAIR generates responses under three reasoning modes for each prompt and uses a composite reward based on correctness, grounding, and cost to train mode selection.

  • Results

    +5.4 on MathVista, +4.8 on Video-Holmes, and +4.8 on Video-MMMU over Qwen2.5-VL, with average token usage reduced to 95 from 417 for always-thinking baselines.

  • Takeaways & Limitations

    Video-FLAIR learns adaptive reasoning without per-query annotations and improves accuracy while substantially reducing token usage across image and video benchmarks.

  • Takeaways & Limitations

    The framework relies on proxy reasoning-quality signals, and verifier calibration may drift on out-of-distribution queries.

Abstract

from arXiv · show

Multimodal queries can require different types of reasoning. Some can be answered via perceptual reasoning, extracting information directly from the visual signal, while others require compositional reasoning that combines observations or deliberative reasoning that evaluates competing hypotheses. However, many existing methods apply a uniform reasoning strategy across queries, leading to unnecessary computation on simple tasks and insufficient reasoning on complex ones. We introduce Video-FLAIR, a training framework that learns to select the appropriate reasoning mode for each query using reinforcement learning. During training, the model generates responses under all three modes for the same prompt, enabling direct comparison. A composite reward compares these responses to favor the most effective one based on correctness, grounding, and cost, while discouraging unsupported or misaligned deliberation. This yields a supervision signal for learning adaptive reasoning without per-query annotations. Video-FLAIR improves accuracy over the Qwen2.5-VL base model by +5.4 on MathVista, +4.8 on Video-Holmes, and +4.8 on Video-MMMU, while reducing average token usage to 95 compared to 417 for always-thinking baselines.

1 Introduction

Video-FLAIR treats multimodal reasoning as query-dependent, distinguishing direct perception, composition, and deliberation instead of applying one strategy uniformly. It learns this adaptive selection through reinforcement learning and within-query comparisons of mode-specific responses.

  • Motivation: Uniform reasoning can waste tokens on simple queries while leaving complex queries without sufficient multi-step or hypothesis-driven reasoning.Long chains may also shift attention away from visual evidence and produce hallucinated intermediate steps.
  • Adaptive reasoning: Video-FLAIR defines PERCEPT, COMPOSE, and DELIBERATE modes for direct evidence mapping, concise observation sequences, and competing-hypothesis evaluation.The modes are designed to match different reasoning demands.
  • Training framework: Video-FLAIR uses reinforcement learning to select an appropriate reasoning mode for each query.The framework learns adaptive reasoning without requiring per-query annotations.
  • Training framework: Mode-structured rollouts generate responses under all three reasoning modes for the same query, enabling direct within-query comparison.A composite reward includes an online verifier to promote grounded reasoning.
  • Results: +5.4 on MathVista, +4.8 on Video-Holmes, and +4.8 on Video-MMMU over Qwen2.5-VL, while average token usage falls to 95 from 417 for always-thinking baselines.These results combine higher accuracy with lower average token usage.

2 Related Work

Prior multimodal reasoning methods commonly use fixed or binary reasoning strategies, while Video-FLAIR compares multiple reasoning structures for the same query. Its training pipeline combines structured rollouts, composite rewards, verifier feedback, and policy updates.

  • Overthinking and adaptive reasoning: Chain-of-thought can use up to 20× more tokens than necessary on simple queries and may cause visual forgetting and hallucinated intermediate steps.Excessive reasoning can also degrade performance when deliberation is unnecessary.
  • Overthinking and adaptive reasoning: Prior adaptive methods often make a binary think/no-think decision, which does not distinguish visual retrieval from tasks requiring reasoning.Examples include confidence-based early exit and trigger-based switching.
  • Reinforcement learning: Each training pair produces K=8 mode-structured rollouts, whose composite rewards guide policy updates and verifier refreshes.The verifier is periodically refreshed using highest- and lowest-reward rollout pairs.
  • Overthinking and adaptive reasoning: ARM2 expands beyond binary switching but relies on fixed SFT-defined templates rather than learning how to select among reasoning formats.This differs from Video-FLAIR's learned mode selection.
  • Reinforcement learning: Reinforcement-learning approaches improve multimodal reasoning but generally continue applying a single reasoning style across queries.Video-FLAIR instead compares alternative strategies on the same prompt.

3 Preliminaries

Video-FLAIR defines an adaptive reasoning space with three modes that differ in how observations are transformed into answers. The modes range from immediate evidence extraction to sequential composition and hypothesis-based deliberation.

  • PERCEPT: PERCEPT (<DIRECT>) extracts a single observation that maps directly to the answer through an immediate, unbranched evidence chain.Its reasoning path is direct rather than multi-step.
  • COMPOSE: COMPOSE (<CONCISE>) combines multiple observations through a bounded linear sequence in which each step builds on the previous one.Its chain is concise because it is bounded, not necessarily short.
  • DELIBERATE: DELIBERATE (<DEEP>) revisits the same evidence to evaluate and reject competing hypotheses.Its depth reflects search over interpretations rather than simply longer chains.

4 VideoFLAIR

Video-FLAIR trains a multimodal policy to compare structured rollouts across reasoning modes, then select the mode with the best correctness, grounding, and cost utility. Its reward combines adaptive selection, verifier feedback, cost control, mode balance, and token-level credit shaping.

  • Framework: Video-FLAIR trains a policy to select among PERCEPT, COMPOSE, and DELIBERATE modes at inference.The framework uses reinforcement learning and mode-structured rollouts to learn adaptive reasoning rather than a fixed strategy.
  • Mode-structured rollouts: For each query, controlled rollouts generate two responses under each reasoning mode, while adaptive slots select a mode without explicit specification.The eight-slot design uses slots 1–6 for controlled modes and slots 7–8 for adaptive selection.
  • Composite reward: Each rollout receives a composite reward combining correctness and validity, cost efficiency, mode balance, adaptive selection, and verifier feedback.The reward includes Rans, Rformat, Rcomply, Rcost, Rbalance, Rselect, and Rverifier.
  • Cost-aware selection: Adaptive length penalties strengthen on easy queries and weaken on difficult queries, discouraging unnecessary deliberation while allowing deeper reasoning when needed.The solved-rollout fraction proxies query difficulty, and per-mode costs penalize deeper modes even within budget.
  • Utility and selection: The selected mode maximizes utility from correctness, grounding, and cost, with cost deciding ties between equally accurate and grounded modes.Grounding resolves near-ties without allowing a grounded incorrect answer to outrank an accurate ungrounded one.
  • Verifier and credit shaping: Verifier feedback measures temporal, spatial, and human alignment, while GDPO and token-level shaping concentrate gradients on evidence-bearing spans.Grounding tags and factual observations receive higher weights, whereas hedging, filler, and repetition receive lower weights.

5 Experiments

Experiments evaluate Video-FLAIR across image and video benchmarks, showing adaptive mode selection improves accuracy, reduces token use, and benefits from structured rollouts and reward components.

  • Evaluation setup: Video-FLAIR is evaluated on six image and five video benchmarks spanning retrieval, hallucination robustness, mathematical and scientific reasoning, spatial understanding, and video knowledge acquisition.
  • Image benchmarks: Video-FLAIR achieves the best performance across all six Qwen2.5-VL image benchmarks while using 74 tokens on average.On Qwen3-VL, it leads on five of six image benchmarks and uses 108 tokens on average.
  • Video benchmarks: 48.0 on Video-Holmes, 35.8 on VSI-Bench, and 56.9 on Video-MMMU are achieved over Qwen2.5-VL while using 59 tokens on average.The reported gains over the base model are +4.8, +4.0, and +4.8, respectively.
  • RL gains over SFT: RL improves over SFT across all benchmarks by using rollout comparisons to learn which reasoning mode each query requires.The largest Qwen3-VL gains are EMMA (+3.9) and MMMU (+5.3), while grounding rewards improve HallusionBench (+1.4) and Video-Holmes (+2.0).
  • Ablations: Removing adaptive, controlled, or mode-selection slots lowers accuracy or increases token usage, showing that structured three-mode rollouts are necessary for cost-aware adaptation.Without adaptive slots, the policy settles on <CONCISE> and token usage rises to 237.
  • Ablations: Three reasoning modes outperform a binary thinking switch because separating <CONCISE> from <DEEP> preserves medium-difficulty and hypothesis-driven reasoning.Collapsing them reduces EMMA to 24.4 and Video-MMMU to 55.1.

6 Conclusion

Video-FLAIR concludes that multimodal models should adapt reasoning structure to each query rather than apply a fixed strategy. Across image and video benchmarks, this approach improves accuracy while reducing token usage, while future work may extend the mode space and permit mode switching mid-response.

  • Conclusion: Video-FLAIR adapts among PERCEPT, COMPOSE, and DELIBERATE modes without requiring per-query annotations.
  • Conclusion: Uniform reasoning fails through hallucinated intermediate steps on perceptual tasks and incorrect interpretations when ambiguous tasks receive insufficient deliberation.
  • Conclusion: Video-FLAIR consistently improves accuracy while substantially reducing token usage compared with always-thinking strategies across image and video benchmarks.
  • Conclusion: Future directions include tool-augmented completions and switching reasoning modes mid-response when initial reasoning is insufficient.

B Ablation Studies

Ablations show that Video-FLAIR’s gains depend on balanced, cost-aware, grounded mode selection, with <DEEP> retained for queries requiring competing-hypothesis elimination.

  • Cross-category gains: Video-FLAIR produces positive gains across all five task categories in both base models, unlike always-thinking and auto-thinking alternatives.Always-thinking models collapse on image perception and spatial reasoning, while auto-thinking recovers perception only partially.
  • Mode distribution: Mode usage tracks cognitive demand: <DIRECT> ranges from 74% on HallusionBench to 30% on SciVideoBench, while <CONCISE> dominates tasks requiring multi-step visual-domain integration.<DEEP> remains below 6% and rises when competing explanations must be ruled out.
  • Reward ablations: Removing the diversity reward collapses selection toward <DIRECT>, reducing Video-Holmes to 44.9 and SciVideoBench to 26.3.HallusionBench is largely unaffected because <DIRECT> is already appropriate there.
  • Reward ablations: Removing cost pressure increases verbose responses and lowers HallusionBench to 69.4 and VSI-Bench to 33.1.Extended chains create visually ungrounded steps on tasks where answers must be read directly from visual input.
  • Reward ablations: Removing grounding from the utility lowers VSI-Bench to 32.9 and Video-Holmes to 45.6 because the utility cannot distinguish grounded from ungrounded correct answers.
  • Mode-space ablation: Removing <DEEP> lowers Video-Holmes to 46.1 and SciVideoBench to 27.8 because <CONCISE> cannot fully replace hypothesis elimination.HallusionBench remains nearly flat because it is dominated by <DIRECT> queries.

D.3 Video Benchmarks

The paper evaluates Video-FLAIR across video benchmarks spanning temporal, spatial, perceptual, domain-specific, and knowledge-acquisition demands. Its training procedure defines structured rollouts, answer-type scoring, format and compliance rewards, and mode-balance incentives.

  • Video benchmark coverage: Video-Holmes tests locating and causally linking visual clues scattered across temporal segments rather than recalling information from one frame.Even the strongest models cited achieve only approximately 45% accuracy.
  • Training and scoring: The training procedure samples K=8 mode-structured rollouts, scores answer correctness and output compliance, and penalizes mode collapse across controlled groups.Its cost reward uses per-mode costs and word budgets, while selection utilities supervise adaptive slots.
  • Video benchmark coverage: Video-TT evaluates correctness and robustness on 1,000 YouTube Shorts with rephrased, wrongly-led, and correctly-led adversarial question variants.Its robustness track probes over-reasoning and confirmation bias.
  • Video benchmark coverage: VSI-Bench uses more than 5,000 indoor-scene-video QA pairs to test spatial measurement, distance estimation, counting, and route planning.The benchmark identifies spatial reasoning as a primary bottleneck for current models.
  • Video benchmark coverage: SciVideoBench contains 1,000 research-level questions across more than 25 domains requiring precise spatiotemporal perception, expert knowledge, and logical reasoning.Its questions span physics, chemistry, biology, and medicine.
  • Video benchmark coverage: Video-MMMU measures knowledge acquisition from 300 expert educational videos through 900 questions aligned to Bloom’s taxonomy.Performance decreases as cognitive demand increases.

E.5.1 Cost Pressure: Scenario Analysis

Video-FLAIR makes reasoning cost sensitive to query difficulty: penalties discourage unnecessary deep reasoning on easy queries while relaxing when controlled rollouts indicate difficulty. Mode selection combines answer quality, grounding, cost, and verifier feedback.

  • Easy-query scenario: On easy queries, high solved-rollout fractions strengthen cost pressure, favoring cheaper reasoning modes when they already solve the task.The difficulty proxy is computed across all K=8 slots.
  • Utility-based selection: Mode utilities combine answer reward, verifier grounding, and mode cost, with ties resolved by a stated tie-breaking rule.Table 8 summarizes representative utility outcomes for DIRECT, CONCISE, and DEEP.
  • Hard-query scenario: On hard multi-hop queries, low solved-rollout fractions relax the adaptive length penalty, allowing deeper reasoning when it is needed.The scenario sets p̂_solved=0.10 and describes the DEEP rollout at 700 tokens.
  • Verifier feedback: Every 100 RL steps, highest- and lowest-reward completions form preference pairs used to refresh the verifier through DPO.Pairs are retained only when their reward margin exceeds 0.15.

G SFT Data

Video-FLAIR uses mode-structured supervised warm-start data to establish DIRECT, CONCISE, and DEEP behaviors before reinforcement learning. The partitions map datasets to near-zero, concise, or deep reasoning demands.

  • Data construction: The SFT corpus converts heterogeneous multimodal QA and reasoning sources into a unified schema with mode-appropriate completions and strict wrappers.The warm start is built on OneThinker and Open-O3-Video traces.
  • Mode partitions: The paper summarizes the three mode partitions and their approximate sample counts in Table 10.
  • Mode partitions: DIRECT contains approximately 39K samples for tracking, segmentation, grounding, and OCR tasks requiring near-zero reasoning.Its traces target no more than 60 think words and emphasize precise localization.
  • Mode partitions: DEEP contains approximately 12.2K samples for causal, logical, intent, disambiguation, movie-QA, and procedural-understanding tasks.These traces can use negative-reasoning style to falsify plausible distractors before answering.

H RL Training Setup

RL training begins from an SFT-initialized policy and uses difficulty-filtered prompts, eight distributed rollouts, adaptive mode slots, and prioritized replay. The identified failure modes show that routing remains imperfect in several regimes.

  • RL data and optimization: Difficulty screening discards questions that are trivially easy or trivially hard for the SFT-initialized policy, retaining prompts with learning signal.The default screening uses 4–8 samples and thresholds based on correct-sample counts.
  • RL data and optimization: Training fine-tunes Qwen3-VL and Qwen2.5-VL with LoRA applied to all linear layers and uses bfloat16 precision with ZeRO-3 sharding.
  • Rollout configuration: Each prompt receives K=8 rollouts with temperature 1.0, top-p 1.0, and a maximum completion length of 1,024 tokens.Generation uses vLLM in server mode.
  • Rollout configuration: Two controlled slots each use DIRECT, CONCISE, and DEEP, while two adaptive slots receive learned mode choices through deterministic slot assignment.Mode instructions are injected at sampling time.
  • Replay and correction: Prioritized replay stores the highest-priority entry per prompt in a 4,096-entry buffer and prioritizes verifier magnitude together with incorrectness.The priority is p = |S_verifier| + (1 − R_ans).
  • Replay and correction: Clipped importance-sampling weights correct replay advantages after the policy changes, limiting destabilizing weights.
  • Failure modes: The analysis identifies mode collapse on easy batches, over-deliberation on difficult queries, and under-invocation of DEEP on Video-Holmes.Approximately 25% of Video-Holmes selections are DIRECT despite the benchmark’s non-direct-lookup design.

K Limitations

Video-FLAIR evaluates reasoning quality through proxy signals and assumes queries can be meaningfully compared across all reasoning modes. Its verifier may drift or share backbone biases, while forced multi-mode rollouts can waste computation on single-depth tasks.

  • Proxy signals, including grounding tags and verifier scores, replace direct human supervision for evaluating reasoning quality.The verifier is periodically refreshed during training, but its calibration may drift on out-of-distribution queries.
  • A shared Qwen3-VL backbone for the policy and verifier may bias verifier judgments.Consistent gains on Qwen2.5-VL, which uses a different policy backbone, suggest this is not the primary improvement source.
  • Slot-based rollouts assume every query can be meaningfully evaluated across all reasoning modes.For intrinsically single-depth tasks such as pure classification, forced CONCISE or DEEP rollouts can waste computation on degenerate traces.

L Qualitative Analysis

Qualitative examples show that mode selection succeeds when reasoning depth matches query structure, but fails through both under-routing complex tasks and over-routing simple lookups. The figures also illustrate the operational prompts and verifier inputs used to structure these modes.

  • Figure 9: Figure 9 shows under-invocation of DEEP for emotional-state assessment requiring posture, gaze, and reaction across multiple frames.The policy routes to DIRECT and produces an incorrect single-frame conclusion, with grounding tags appearing inside the DIRECT output.
  • Figure 10: Figure 10 shows CONCISE failing on rover-autonomy assessment because it chains observations without falsifying competing interpretations.The query requires evaluating controller presence, motion smoothness, and gaze tracking; DEEP is required, whereas wheel counting is a DIRECT lookup.
  • Figure 11: Figure 11 shows DEEP correctly disambiguating cube causality, while DIRECT misses one object on a temporal counting task.The causality analysis uses timestamped bounding boxes; the counting query requires a cumulative scan across the video, with the correct total being 4 objects.
  • Figure 12: Figure 12 shows DEEP correctly eliminating three distress hypotheses, while CONCISE over-routes a single-frame object-identification lookup.For the lookup, DIRECT and CONCISE produce approximately equal answers, but CONCISE has higher cost and should lose on utility.
  • Adaptive inference: ADAPTIVE inference asks the model to choose the cheapest reliable style among DIRECT, CONCISE, and DEEP, using verifier inputs that include the question, completion, ground truth, and media context.The verifier evaluates correctness, grounding, human alignment, and whether the selected reasoning mode is appropriate.
Loading 2608.26495v1…