PaperlayerView exact revision on arXiv

Source-linked AI summary

StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding

Keming Wu, Baoyi Wang, Kaichen Zhang, Xiang An, Zuhao Yang, Sudong Wang, Haowei Zhu, Tingxuan Huang, Hongcheng Gao, Bin Wang

arXiv:2608.16320v1cs.CV

TL;DR

Streaming video models must answer from causal prefixes, while existing inference machinery adds complexity without clear necessity. StreamOPD isolates post-training under a memory-free recent-window protocol and raises the 4B student to 83.9% on StreamingBench, within 0.3 points of the 9B teacher.

  • Problem

    Can post-training close the small-to-large model gap in streaming video while keeping a simple memory-free inference path unchanged?

  • Method

    StreamOPD combines verifiable streaming-video data, thinking-mode OPD training, instruct-mode deployment, and optional cue-gated teacher reweighting under fixed recent-window inference.

  • Results

    83.9% on StreamingBench places the 4B student within 0.3 points of the 9B teacher, while StreamOPD improves OVO-Bench excluding HLD by 9.1 points under unchanged inference.

  • Takeaways & Limitations

    Post-training alone can substantially improve streaming-video understanding without adding inference-time memory, retrieval, compression, or online reasoning modules.

  • Takeaways & Limitations

    ST-CueGate evaluates one group-relative likelihood gate, leaving alternative aggregations and absolute or token-selective gating unexplored.

Abstract

from arXiv · show

Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to \textsc{StreamOPD}, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from $77.9\%$ to $83.9\%$---within $0.3$ points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by $9.1$ points under unchanged inference. As a teacher-privilege extension, \emph{Spatio-Temporal CueGate (ST-CueGate)} aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches $71.9\%$ on OVO-Bench (excluding HLD) and $64.9\%$ on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to $57.0\%$, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.

Introduction · Method

StreamOPD shows that post-training can substantially improve streaming video understanding under a fixed memory-free recent-window protocol, with thinking-mode OPD enabling stable training and instruct-mode deployment. Its ST-CueGate extension uses teacher-side cue sensitivity to selectively reweight distillation without changing student inference.

  • Introduction: Only both-thinking teacher/student training converges in the tested mode pairings and reaches 83.9% on StreamingBench when evaluated in instruct mode.Both-instruct and teacher-thinking/student-instruct runs collapse early with very short responses and substantially larger gradient norms.
  • Introduction: 83.9% on StreamingBench raises the Qwen3.5-4B student from 77.9% to within 0.3 points of the Qwen3.5-9B teacher, while OVO-Bench excluding HLD improves by 9.1 points under unchanged inference.These gains establish the core empirical value of StreamOPD while preserving the fixed deployment protocol.
  • Method: StreamOPD combines a reproducible 25k verifiable-data pipeline, thinking-mode OPD training, instruct-mode deployment, and a fixed memory-free recent-window evaluation protocol.The recipe responds to failures of teacher-free GRPO and OPD configurations involving instruct-mode training.
  • Introduction: The method fixes inference to the most recent four frames at 1 fps, excluding memory banks, retrieval, and compression so post-training is evaluated without architectural changes.This protocol follows SimpleStream’s finding that a training-free recent-window baseline can match more complex systems, reaching 67.7% on OVO-Bench and 80.6% on StreamingBench.
  • Method: ST-CueGate compares cue and no-cue teacher likelihoods on each student response, aggregates their token-level contrast, and applies a group-relative scalar weight to OPD advantages.The cue is teacher-only, while the student prompt and inference protocol remain unchanged; above-group-mean responses are up-weighted and below-mean responses down-weighted.
  • Method: 56.5% of cue-contrast tokens are near zero, 23.1% are negative, and the top 20% of positive tokens contain 82% of positive mass, motivating selective rather than uniform cue conditioning.ST-CueGate uses a length-normalized response score standardized within a minibatch or sibling-rollout comparison group.
  • Method: The training set contains approximately 25k verifiable MCQ, binary, and counting questions with automatically screened spatio-temporal cues generated from complete clips after answer options are removed.Rule-based screening, regeneration, and deterministic judging accept approximately 96.6% of candidate cues, which are supplied only to the teacher.

Experiments

Under fixed recent-window inference, StreamOPD substantially improves the 4B student through thinking-mode OPD, while ST-CueGate further strengthens results and self-distillation preserves most gains without requiring a larger teacher. Experiments also identify response-level, moderately weighted cue gating as most stable, while GRPO suffers severe format drift.

  • Training–Deployment Mode Decoupling and Stability: 82.4% under the last-letter parser versus 35.1% under the deployed first-answer parser reveals GRPO’s format drift, while both-thinking OPD training converges and remains concise.GRPO responses grow more than tenfold in median length, whereas distilled models remain parser-invariant; teacher scale alone also does not guarantee better on-policy supervision.
  • StreamOPD Core Recipe Under Fixed Recent-Window Inference: 83.9% on StreamingBench and +9.1 points on OVO-Bench excluding HLD show that standard OPD closes most of the 4B–9B gap under unchanged inference.Gains span real-time perception and backward tracing, with the largest subtask improvements on STU (+20.7) and OCR (+13.4).
  • Teacher-Privilege Extension: ST-CueGate: 83.35% on StreamingBench and 67.35% on OVO-Bench excluding HLD show that self-distilled ST-CueGate retains most gains, while its 57.0% HLD score exceeds the untrained student and 9B teacher.The frozen 4B same-backbone teacher makes ST-CueGate compatible with self-distillation rather than dependent on teacher scale.
  • ST-CueGate Ablations: Response-level n=4 gating performs best on StreamingBench, OVO-Bench, and LongVideoBench, while n=1 slightly favors Video-MME; token-wise gating loses 3.2 and 1.4 points on the former two.Aggregating cue contrasts at the response level is more stable than using individual token contrasts as independent weights.
  • ST-CueGate Ablations: αg = 0.5 gives the highest result on all four benchmarks, whereas αg = 1.0 consistently hurts performance.The likelihood contrast is most useful as a moderate ranking signal rather than an aggressively amplified weight.

Related Work

Streaming video understanding operates on causal prefixes, with prior systems using memory, retrieval, or compression despite recent-window baselines matching heavier designs. OPD and privileged-information methods motivate ST-CueGate’s fixed-trajectory, cue-removal contrastive reweighting.

  • Streaming Video Understanding: Streaming models process causal video prefixes and commonly use memory, retrieval, or KV-cache compression, while StreamingBench and OVO-Bench evaluate real-time and historical reasoning.Offline video MLLMs instead process complete clips. Relevant systems include Chen et al. 2024; Zhang et al. 2024a; Qian et al. 2025; Yao et al. 2025; Zeng et al. 2026; Zhang et al. 2026; Xiao et al. 2024.
  • Streaming Video Understanding: SimpleStream (Shen et al. 2026) shows that a recent-window baseline can match heavier streaming designs, whereas ThinkStream (Liu et al. 2026) combines memory with online reasoning.
  • On-Policy Distillation and Teacher Privilege: OPD scores student-generated trajectories rather than fixed teacher targets and has extended to text reasoning, multimodal pre-alignment, and temporal video grounding.Classical distillation uses teacher targets on fixed data (Hinton, Vinyals, and Dean 2015), whereas OPD is associated with Agarwal et al. 2024; Gu et al. 2024; Zhao et al. 2026; Wang et al. 2026; Li et al. 2026.
  • On-Policy Distillation and Teacher Privilege: ST-CueGate builds on privileged-information and visual-cue gating work by contrasting teacher likelihoods with a cue-removal reference while holding frames, question, and student trajectory fixed.Privileged-information methods provide training-only side information (Vapnik and Izmailov 2015; Lopez-Paz et al. 2015), while related visual-cue methods include Tian et al. 2026 and Sun et al. 2026.

Conclusion

Under a fixed memory-free recent-window protocol, the study isolates post-training from added inference machinery and identifies failure modes that motivate the StreamOPD recipe. The recipe combines a 25k verifiable-data pipeline with thinking-mode training and instruct-mode inference.

  • Core recipe: StreamOPD combines a 25k verifiable-data pipeline with thinking-mode training and instruct-mode inference under a fixed memory-free recent-window protocol.This setup isolates post-training effects from added inference machinery.
  • Failure modes: Teacher-free GRPO drifts toward verbose responses that are incompatible with deployment.
  • Failure modes: The tested OPD configurations involving instruct-mode training collapse, motivating the thinking-train/instruct-infer configuration.

Implementation and Data Details · Recent-Window vs. Dense-Uniform Context · Prompt Templates

The study uses reproducible video-OPD infrastructure, compares recent-window and dense-uniform causal context, and specifies leakage-controlled cue and evaluation prompts. Dense context improves historical recall but harms recent-event perception, supporting the deployment-aligned recent-window protocol.

  • Implementation and Data Details: Training runs use HybridFlow-style RL, vLLM rollouts, NVIDIA H200 GPUs, response-only teacher log-probabilities, corrupt-video tolerance, and verifiable rewards for MCQ, binary, and counting tasks.The customized pipeline aligns teacher and student despite differing prompt lengths and skips unreadable clips.
  • Implementation and Data Details: Settings are selected one axis at a time using only the held-out validation aggregate, never downstream benchmark scores, with Table 7 listing explored values and reported choices.The main ST-CueGate configuration fixes gate range [wmin, wmax]=[0, 2], γ=1, and four epochs.
  • Implementation and Data Details: All checkpoints use instruct-mode greedy decoding, with recent-window evaluation for StreamingBench and OVO-Bench and standard protocols for Video-MME and LongVideoBench.OVO-Bench reports full macro, the score excluding HLD, and raw HLD because HLD measures hallucination robustness rather than episodic event recall.
  • Implementation and Data Details: Each trained configuration uses seeds 42, 43, and 44; reported entries average three selected models, while standard deviations and paired significance tests are unavailable.Greedy evaluation removes sampling noise but not training or checkpoint variance.
  • Recent-Window vs. Dense-Uniform Context: Dense-uniform context improves historical recall but reduces recent-event performance: ASI rises 12.17 and EPM 7.48, while realtime macro falls 10.50, HLD 15.05, and StreamingBench 3.04 points.The comparison uses one fixed ST-CueGate model; dense-uniform samples up to 32 visible-prefix frames, whereas Recent-4 retains the latest four.
  • Prompt Templates: The student prompt remains identical in training and deployment and contains no cue, while the teacher receives the student content plus explicit visual-evidence instructions.ST-CueGate’s no-cue teacher pass reuses the original student prompt; fallback samples satisfy Δ=0 and use a neutral gate.
  • Prompt Templates: Cue generators produce pure spatio-temporal pointers without answers, using stricter binary-item rules; flagged cues are regenerated and automatically judged SAFE or LEAK.The judge permits location or timing references but rejects answer attributes, outcomes, identities, counts, yes/no content, and quoted on-screen text.
  • Prompt Templates: StreamingBench and OVO-Bench use direct-answer templates, whereas Video-MME and LongVideoBench retain their unmodified standard-protocol prompts.StreamingBench requires only the best option letter, and OVO-Bench’s backward-tracing and real-time subtasks use a shorter equivalent template.

Cue Construction and Screening

The cue pipeline generates short spatio-temporal pointers while screening answer leakage through automatic rules and manual audits. Accepted cues are predominantly non-leaking, though the audit estimates precision rather than false-rejection rates.

  • Offline Cue Generation: Cues are generated as single sentences of at most 25 words that point to when or where to inspect without stating answers or attributes.Qwen3.5-27B generates one spatio-temporal pointer per training sample in the implementation.
  • Leakage Taxonomy and Screening: The screening taxonomy covers direct answers, semantic paraphrases, discriminative attributes, outcome clues, and cues that eliminate alternatives.Rule-based checks detect option letters, binary assertions, explicit counts, quoted answer text, and lexical overlap; failed candidates are regenerated up to three times.
  • Automatic Screening: 96.6% of 25,118 samples were automatically accepted, while 3.4% fell back to the original teacher prompt.The pipeline complements same-model automatic judging with a manual audit of accepted cues.
  • Manual Audit: 7/300 accepted cues leaked under the first two categories, yielding 2.3% residual leakage and approximately 97.7% estimated acceptance precision.The audit’s end-to-end non-leaking fraction is approximately 94%, but it does not estimate false rejection among fallback samples.
  • Representative Accepted Cues: Representative accepted cues localize moments, subjects, or regions without naming colors, counts, or answer options, preserving the need for visual inspection.Table 10 samples two cues per answer format from the released training pool.

Answer Parsing and Verifiable Reward

Typed answer parsing assigns exact-match rewards for GRPO diagnostics, while OPD never uses the reward. Parser direction creates a format-drift gap for rationale-heavy GRPO outputs, whereas distilled models’ bare option letters make deployed and training parsers agree.

  • Reward definition: Exact-match reward is 1 after type-specific extraction and 0 otherwise; this rule-based reward drives GRPO diagnostics but never enters the OPD objective.Ground truths are automatically typed as A–H, yes/no, or non-negative integer, with a dedicated extractor for each format.
  • Parser behavior: The training parser takes the last qualifying multiple-choice letter, whereas the deployed streaming parser takes the first answer token.Multiple-choice extraction prioritizes the last explicit answer cue or otherwise the last standalone uppercase letter; binary and counting extraction use the first qualifying token or integer.
  • Format-drift diagnostic: Long rationales before answers score well under training parsing but poorly under deployment parsing, producing the GRPO format-drift gap.Distilled models emit bare option letters, so both parsers agree.

Training Algorithm

ST-CueGate training samples student thinking-mode responses, replays them with the teacher under cue and original prompts, and uses a group-relative gate to reweight token-level OPD advantages. The gate is a relative response measure whose scalar weight applies across all response tokens, including reasoning, answer, punctuation, and formatting.

  • Gate Interpretation: The gate is group-relative rather than an absolute cue-benefit test, so an above-mean response can receive w_i > 1 even when every sibling response is negative.This interpretation follows from standardizing g_i within each prompt’s sibling rollout group before setting the clipped weight.
  • Gate Interpretation: The response mean mixes reasoning, answer, punctuation, and formatting tokens, and its scalar reweights every token-level OPD advantage.The cited passage notes that alternative absolute, positive-clipped, positive-token-only, and answer-token aggregations were not evaluated.
  • ST-CueGate Training: ST-CueGate replays each on-policy thinking-mode response with the teacher under cue and original prompts, then standardizes responses within sibling rollout groups to compute a clipped gate.The procedure forms a scalar weight from the group-relative response and applies it to token-level OPD advantages.

Objective Derivations and Policy-Gradient · View

The derivations show that OPD is a dense-reward policy-gradient method using an unbiased sampled-token estimator, while extrapolated targets can point toward ungrounded, unrealizable distributions. ST-CueGate instead uses a length-normalized cue/no-cue likelihood ratio to mean-preservingly reweight sibling rollouts, with clipping behavior determined by group size and gate strength.

  • View: The single-sampled-token estimator is unbiased for the per-token reverse-KL term, and all reported runs use the signed k(1) estimator because it supports the advantage interpretation.Full-vocabulary summation is expensive for long multimodal prompts; lower-variance k(2) and k(3) estimators are available but non-negative by construction.
  • View: OPD performs gradient ascent on the dense per-token reward r_i,t = τ_i,t − s_i,t, which is positive when the teacher favors the sampled token and vanishes when student and teacher match.The reward is self-annihilating, unlike GRPO’s response-level constant advantage, which can vanish when every rollout receives the same outcome.
  • View: For λ > 1, ExOPD targets b_i,t + λ(τ_i,t − b_i,t), which generally is not the log-density of any normalized teacher policy and therefore pushes toward an unrealizable distribution.Because the extrapolation direction is not tied to an observable cue in the training clip, this ungrounded signal is consistent with the reported instability.
  • View: AD-ExOPD decomposes teacher improvement into frame and capacity axes using a dense-frame base reference, with separate gains; λ_t = 1.5 and λ_c = 1.0 were used for the reported experiment.Setting λ_t = λ_c recovers ExOPD, while the chosen gains reflect the student’s access to additional frames but not 9B capacity.
  • View: ST-CueGate’s sequence-level contrast is a length-normalized teacher likelihood ratio between cue and no-cue contexts for the same realized response, rather than an information-theoretic mutual-information quantity.The two conditionals come from prompting one network in two ways and need not define a consistent joint distribution over cues and responses.
  • View: The gate is mean-preserving within each sibling group, redistributing weight without changing the group’s total contribution; α_g controls spread, with |e_g| ≤ √n − 1 and attainable weights [1 − α_g√n − 1, 1 + α_g√n − 1].For n = 4 and α_g = 0.5, the attainable range lies strictly inside the clipping interval, so clipping cannot bind.
  • View: Increasing α_g to 1.0 at n = 4 or increasing group size to n = 8 at α_g = 0.5 makes both clipping boundaries reachable, and both settings lose accuracy in the corresponding ablations.The paper presents this as a consistency observation rather than an identified causal explanation.
  • View: OPD, ungrounded controls, and contrastive gates share a group-relative reweighting mechanism; their substantive difference is the reference used to generate w_i.The references include constant weights for OPD, student–teacher disagreement for DAD, positive/degraded visual views for V-Zero, and cue/no-cue teacher contexts for ST-CueGate.

Data Construction and Statistics

The training set comprises short, predominantly temporal streaming-video QA items built from public instruction data, then filtered to remove examples that are easy for both student and teacher models. The resulting 8,343-item set includes a small teacher-underperformance subset largely attributable to response-format deviations.

  • Data Construction and Statistics: 71.8% of questions require temporal reasoning, while videos have a median duration of ∼5.5s and 89% are shorter than 15s.The question types are 69.7% multiple-choice, 25.3% binary, and 5.0% counting.
  • Data Construction and Statistics: 8,343 items remain after pass-rate filtering removes samples already easy for both models.Student and teacher outputs are sampled eight times each at temperature 0.7 to estimate pass rates.
  • Data Construction and Statistics: 40.7% of the 4.6% of samples where the teacher underperforms the student reflect response-format deviations rather than genuine capability gaps.The filtering procedure samples both models eight times at temperature 0.7 before estimating per-sample pass rates.

Cue-Conditional Contrast Analysis · Negative Results: A Map of the Design Space

Cue-conditional teacher scores remain sparse, heavy-tailed, and concentrated across training checkpoints, supporting ST-CueGate’s design. Across alternative reward, distillation-mode, frame-budget, and teacher-free RL variants, no method consistently improves over standard OPD under the fixed protocol.

  • Cue-Conditional Contrast Analysis: The inactive-token fraction remains 52.4–56.5%, while the top 20% of positive-∆ tokens account for 80–82% of positive mass across three checkpoints.The aligned distributions at the base model, step 700, and step 2100 show that the sparse, concentrated cue-conditional structure is not limited to initialization.
  • Negative Results: A Map of the Design Space: The fixed evaluation protocol uses the identical 25k backbone, optimizer, and evaluation setup across documented variants unless noted.These negative results delineate the tested boundary of post-training and teacher-privilege design.
  • Negative Results: A Map of the Design Space: AD-ExOPD validation oscillates between 50–65%, peaks near 65% at step 1600, and never approaches standard OPD’s approximately 70%.Using separate frame and capacity extrapolation gains also harms early convergence.
  • Negative Results: A Map of the Design Space: DAD reaches 84.27% on StreamingBench but only 68.10% on OVO-Bench (excl. HLD) and 62.93% on Video-MME, providing no consistent OPD improvement.Its response-disagreement weighting is observation-agnostic, unlike the teacher-side grounded privilege retained by the main method.
  • Negative Results: A Map of the Design Space: Mode-mismatched OPD collapses: teacher-thinking/student-instruct validation falls from 69.3% to below 2%, while both-instruct drops from 69.9% to 23.6% by step 100.The recipe therefore trains teacher and student both in thinking mode.
  • Negative Results: A Map of the Design Space: Teacher-free GRPO diverges early in instruct mode, whereas thinking-mode GRPO plateaus near the untrained student under the format-drift diagnostics.The comparison tracks response-length growth and sensitivity to whether the evaluator reads the first or last answer.
  • Negative Results: A Map of the Design Space: For V-Zero adaptation, ST-CueGate uses shared frames, question, and realized response while differing only in cue presence, and the reported negative evidence view temporally shuffles the video.A blacked-out negative view was also implemented, but the shuffled variant was reported because it preserves appearance statistics while destroying the relevant temporal structure.
  • Negative Results: A Map of the Design Space: AFD v2/v3 plateau near 60% validation, never exceed OPD’s 70.6% peak, and trail OPD by about 10 points at matched steps.Increasing the teacher’s frame budget or making its frames a strict superset of the student’s does not improve distillation on the same 25k data.

Broader Impact and Deployment · Considerations

Streaming-video deployment requires privacy and energy considerations, while distilled models may answer unanswerable queries more readily, making explicit abstention evaluation safety-relevant. The analysis also finds that OPD preserves concise, parser-invariant responses, unlike pure RL’s response-format drift.

  • Considerations: Distilled streaming models can become more prone to answering unanswerable queries, so practitioners should evaluate abstention explicitly before deployment.This HLD degradation is safety-relevant but distinct from streaming event recall.
  • Considerations: Public benchmark and instruction data, together with compute-efficient models, lower deployment energy costs, although live camera feeds raise privacy concerns.
  • Considerations: The identified HLD degradation is separate from streaming event recall but remains a deployment-relevant safety concern.The distinction means streaming performance alone does not establish safe abstention behavior.
  • Considerations: 80–82% of total positive cue-conditional likelihood-ratio mass remains concentrated in the top 20% of positive-∆tokens across checkpoints.The cue-versus-no-cue ∆t distributions and positive-score concentration remain stable from the base model through steps 700 and 2100.
  • Considerations: Pure RL outputs grow from concise answers into long rationales, whereas OPD remains concise.This diagnoses response-format drift under sparse reward.
  • Considerations: Pure RL becomes sensitive to whether the evaluator reads the first or last answer, while concise OPD remains parser-invariant.
Loading 2608.16320v1…