STRAND: Benchmarking and Improving Object-Centric Spatio-Temporal Monitoring in Video Large Language Models

Grading only the final answer records that a model was wrong, not which fact it got wrong. STRAND pairs every target question with the atomic sub-questions it rests on, and credits a model only when the answer and all of its supporting facts are correct.

Thong Nguyen†, Tri Cao, Khoi Le, Cong-Duy Nguyen, Quynh Vo, See-Kiong Ng, Bryan Hooi Kuen-Yew

National University of Singapore  ·  Centre for AI Research, VinUniversity

†Corresponding author — thong.nguyen@u.nus.edu

Top: a single answer records only that the model was wrong, whereas STRAND's sub-questions name the facts the answer rests on. Bottom: existing MLLMs answer directly from raw frames, while our framework answers from object trajectories built by chunk-wise state extraction and temporal aggregation.
Existing benchmarks and MLLM architectures versus ours. Top: a single answer records only that a model is wrong, whereas sub-questions name the facts it rests on. In the football clip, #22 shoots and is blocked before #10 shoots, so “does player #1 block the shot of player #10?” is settled only by the order of the two shots. Bottom: we answer from object trajectories rather than raw frames alone.

The primary metric

Faithful Accuracy

A model earns credit for a target question only if it answers that question and every supporting sub-question correctly. The denominator is all 977 targets, so a model cannot raise its score by being selectively consistent on the subset it happens to get right.

Ours (Gemini-3-Flash) 59.4
Ours (Qwen3-VL-235B) 54.2
Qwen3-VL-32B Think 46.1
Qwen3.5-27B 45.5
Gemini-3.1-Pro 38.9
Claude-4.6-Sonnet 30.2
GPT-5 20.0
Caption-only 31.2
Question-only 21.4
Majority class 19.1

Faithful Accuracy (%) on STRAND. The three hatched rows are blind references: a majority-class predictor, a question-only predictor that never sees the video, and a caption-only predictor. Scores should be read against these rather than against 50%.

The four metrics

Faithful Accuracy  AfaithPrimary
The fraction of all targets for which the target answer and every supporting sub-question are correct. Fixed denominator, monotone in both target and sub-question correctness, and zero for a model that answers nothing — so it compares directly across models.
Target Accuracy  Atarget
Standard accuracy on the compositional target questions.
Sub-question Accuracy  Asub
Accuracy over all 2,516 atomic sub-questions — foundational perception, independent of the target.
Conditional Consistency  AconsReference only
Among targets a model answers correctly, the fraction of its sub-questions that are also correct. Because membership of that set depends on the model, two models are scored on different subsets. It is maximised by low-recall selectivity and does not fall when a model misses a target outright, so we report it for reference and state every consistency claim in terms of Afaith.
Why this matters. Cosmos-Reason2-8B posts the best Acons on the benchmark (83.7) while answering 22.4 points fewer targets correctly than our Gemini instantiation. Under a conditional metric that looks like consistency; under Afaith it is 41.5 against 59.4.

The benchmark

STRAND

Rather than generating question–answer pairs from raw video, STRAND begins from a human-verified world model of each video — objects, their states, their relations, and the moments these change. Targets are composed from that record and paired with the atomic sub-questions they depend on.

88
videos, 183 s on average
977
target questions
2,516
sub-questions (2.58 per target)
50
trained annotators, κ = 0.88
100%
yes/no items — graded without a judge model

Videos span sports, egocentric, surveillance and instructional domains, at roughly 22 videos per domain. Because that is thin, we report domain breakdowns as descriptive only and make no claim that the ordering of models is stable within any one domain.

Five-panel overview of STRAND construction: raw video input, expert-driven fact annotation with temporal state anchoring, entity registration, fact extraction and cross-time identity linking, an object-centric structured representation, automatic QA generation and decomposition, and human verification.
Construction pipeline. Annotators anchor timestamps, register entities, extract typed facts and link identities across time. Target questions and sub-questions are composed from that record and then verified against the video.

Annotation

  1. Temporal state anchoring

    Annotators anchor a discrete set of key timestamps at salient interactions, state transitions and occlusions.

  2. Entity registration

    Task-relevant entities receive canonical, visually grounded identifiers such as player_24. An entity is registered only if it is re-identifiable — a human can point to the same entity at two separated timestamps from visual evidence alone — and dynamic, meaning it takes part in at least one state change or interaction. Static background elements are excluded.

  3. Fact extraction

    Observable states and relations are recorded at each timestamp as typed tuples such as (t, o₁, passing_to, o₂), drawn from a three-part schema of attribute, action and relation. No bounding boxes: the failure mode STRAND targets is symbolic binding drift, not localisation error.

  4. Cross-time identity linking

    Entities are traced across the timeline into trajectories. Observation windows need not be contiguous — when an entity is occluded, leaves the frame and re-enters, the gap is recorded explicitly and an identity link is asserted across it only when re-identification is supported by visual evidence such as a jersey number. Otherwise the two segments stay distinct entities. Facts an annotator can infer but not observe are never recorded.

Question generation and verification

Llama-3.1-70B-Instruct organises the annotated facts into reasoning chains, composes a target question from a chain, and decomposes it into the atomic sub-questions the target draws on. The generator operates only on the symbolic fact tuples and never sees the video, so any error it introduces is one of phrasing or unsupported inference rather than invented visual detail. It is also disjoint from every evaluated model family, so no evaluated system shares pretraining lineage with the model that phrases its questions.

Annotators then review every generated item alongside the source video and check that the language is clear and natural, that the sub-questions are jointly sufficient to answer the target and that each probes a fact the target actually draws on, and that every answer is unambiguously grounded in visual evidence. Items carrying unsupported assumptions or unreliable identity bindings are discarded. An alternative sufficient fact set exists for 5.8% of targets.

Can the questions be shortcut?

Single-frame solvability audit. Restricting four frontier models to a single centre frame drops target accuracy to 28.2–32.7% and Faithful Accuracy below 15%. The last two rows receive no visual input at all. This establishes that STRAND items are not solvable from static cues or language priors for the models audited.
Model Full video Restricted input
AfaithAtarget AfaithAtarget
Single centre frame
Gemini-3.1-Pro38.961.414.131.4
Gemini-3-Flash28.748.712.529.8
GPT-520.032.510.228.2
Qwen3.5-27B45.568.413.832.7
No visual input
Question-only (Llama-3.1-70B)——21.452.6
Caption-only——31.259.8

Restricted to the ordering sub-questions, the question-only predictor scores 49.3% — no better than guessing. The order of events has to be read off the video. Of the 382 temporally constrained targets, 361 (94.5%) carry an ordering sub-question.

Main results

STRAND is hard for frontier models

No end-to-end model exceeds 68.4% target accuracy, and several fall below 50%. Three systems score below the 19.1% majority-class Faithful Accuracy: without reliable temporal tracking they cannot ground a target answer in the facts it rests on.

Results on STRAND (%). Best per column in bold, second-best in grey. Afaith is the primary metric; Acons is reported for reference only. Four further open-weight models and the budget-aware selection pipelines appear in the paper's appendix.
Method Afaithprimary ↑ Atarget↑ Asub↑ Aconsreference
Ours
Ours (Gemini-3-Flash)59.477.176.381.3
Ours (Qwen3-VL-235B)54.274.072.678.1
Proprietary end-to-end MLLMs
Gemini-3.1-Pro38.961.464.969.2
Gemini-3-Flash28.748.761.769.9
Gemma-4-27B35.855.560.072.7
GPT-520.032.543.270.5
Claude-4.6-Sonnet30.248.362.973.2
Open-weight end-to-end MLLMs
InternVL3-78B29.749.966.171.0
Qwen3-VL-32B Think46.165.070.678.1
Qwen3.5-27B45.568.466.773.3
Cosmos-Reason2-8B41.554.774.783.7
VideoRFT-7B16.438.655.150.8
Video-R1-7B4.324.819.621.9
Structured pipelines
VideoMind-7B18.950.337.943.2
TraveLER22.050.641.849.3
SeViLA26.151.459.656.8

The Gemini-3-Flash instantiation reaches 59.4 Afaith, ahead of Gemini-3.1-Pro by 20.5 points, with gains of 15.7, 11.4 and 12.1 on the remaining three metrics. That margin sets the scale of the effect; it does not isolate its cause, which the matched-budget comparisons and ablations address.

Scope of the significance test. Significance is assessed by paired bootstrap over STRAND target accuracy, with 10,000 resamples clustered at the video level. Each instantiation improves significantly over the strongest end-to-end model of its own family (p < 0.005). Against Qwen3.5-27B — the strongest open-weight model by target accuracy — the Gemini instantiation gains [+6.8, +10.6] points at 95% confidence. At 88 videos the standard error near 70% target accuracy is 1.5 points, so aggregate differences above 10 points sit well outside sampling noise, while ablation differences below 2.0 points are reported as ties.

What remains hard

Temporal ordering is the bottleneck

Splitting sub-question accuracy by probe type locates the residual errors. Both instantiations identify events far more reliably than they order them, and the gap is roughly 18 points on either backbone.

Sub-question accuracy (%) split by probe type. Both instantiations identify events far more reliably than they order them.
SystemIndividual factsOrder of two eventsGap
Ours (Gemini-3-Flash)78.861.4−17.4
Ours (Qwen3-VL-235B)75.456.2−19.2
Sub-question error types. Shares of each model's own incorrect sub-questions; rows sum to 100. These are shares, not absolute error rates.
ModelObject halluc.State misattr.Temporal order
Ours (Gemini-3-Flash)12.434.153.5
Gemini-3.1-Pro38.236.525.3
Qwen3.5-27B33.738.427.9

Our framework suppresses object hallucination and state misattribution, so ordering becomes the dominant remaining failure at 53.5% of its errors — a 19.4-point margin over the next category. Because these are shares of each model's own errors, the shift reflects what is left once the other modes are reduced, and it matches the ordering gap above. Ordering is the capability STRAND isolates, and the one that remains open.

By question type

Faithful Accuracy (%) across the four target categories, which partition the 977 targets. Identity tracking and action sequencing are the hardest categories for every model evaluated.
Model State changen = 200 Identity trackingn = 277 Action sequencingn = 300 Multi-object interactionn = 200
Ours (Gemini-3-Flash)70.049.852.073.0
Gemini-3.1-Pro50.031.132.748.0
Qwen3.5-27B58.037.938.055.0

Our framework

Answering from object trajectories

Because STRAND credits an answer only when the facts supporting it are correct as well, we build an answering system that represents those facts explicitly. A video becomes a set of object trajectories, each recording how one object's state changes over time. The representation is built once per video and reused by every question about it.

Seven-stage pipeline: raw video input, temporal chunking, chunk-wise state extraction, temporal aggregation and trajectory memory run once per video; query-based trajectory retrieval and trajectory-guided answering run once per question.
Framework overview. Stages 1–5 run once per video. Stages 6–7 run once per question.
  1. Once per video · chunk-wise state extraction

    The video is partitioned into disjoint 15-second chunks and 60 frames are sampled from each. A VLM-based extractor processes every chunk independently, so extraction parallelises. A state is relational rather than purely attributive: it carries the predicate the object participates in together with its arguments.

  2. Once per video · temporal aggregation

    Chunk-level observations contain duplicate objects, noisy states and identities fragmented across time. A deterministic symbolic aggregator consolidates them — it issues no model call, and its parameters (Δtmax, τconf) are fixed hyperparameters, not learned weights. Four steps: similarity scoring over visual attributes and spatial proximity; constraint filtering on temporal proximity and a confidence threshold; bipartite conflict resolution prioritised by temporal adjacency, giving a one-to-one identity mapping; and trajectory generation, in which observations are merged and sorted by timestamp and anything that matches nothing survives as a singleton rather than being discarded.

  3. Per question · trajectory retrieval

    Not every trajectory bears on a given question, so a retriever selects the query-relevant subset. For the stronger backbone this acts mainly as a context-length control rather than a source of accuracy; the weaker backbone is markedly more sensitive to it.

  4. Per question · trajectory-guided answering

    The answer generator receives the retrieved trajectories, the question, and a fixed uniformly sampled 64-frame budget that is independent of the question — so among the video-derived inputs only the retrieved subset varies per query. Operating over structured trajectories rather than raw frames lets this stage address object identities, state transitions and temporal relations directly.

Instantiations

  • Ours (Gemini-3-Flash) — gemini-3-flash-preview for extraction, retrieval and generation.
  • Ours (Qwen3-VL-235B) — qwen3-vl-235b-a22b-thinking for extraction, qwen3-235b-a22b for retrieval, qwen3.5-27b for generation.

Unless ablated, both use 15-second chunks, 60 frames per chunk, 64 frames at the answerer, and identity linking. Every prompt is given verbatim in the paper's appendix, for our framework and for every baseline.

Evidence for the design

What the gain actually rests on

A harness brings extra compute as well as a different representation. The leaderboard above does not separate the two, so we hold the backbone fixed and equalise one resource at a time. These comparisons, not the end-to-end rows, are what the design claim rests on.

Budget-matched comparisons on Faithful Accuracy (%). All configurations use the Qwen3-VL backbone, so only the representation changes within each control.
ConstraintEnd-to-end baselineOursΔ
Equal total frame budget (64 frames)41.554.2+12.7
Equal number of model calls (single pass)39.849.3+9.5
Equal token budget45.152.8+7.7

Component ablations

Target accuracy (%) under component ablations, against the base configuration of each backbone. Differences below 2.0 points are indistinguishable from zero under the video-clustered bootstrap and are reported as ties.
ConfigurationGemini-3-FlashQwen3-VL-235B
Base configuration77.174.0
No retrieval filter (full timeline)77.0tie64.6−9.4
Random retrieval filter (top-k)76.5tie68.0−6.0
LLM summarisation instead of symbolic aggregation76.7tie62.7−11.3
Answerer lowered to 8 frames76.5tie63.6−10.4
Dense prose captions instead of structured states72.4−4.761.2−12.8
Shuffled trajectory order65.3−11.854.1−19.9
Text-only answering (no frames at the answerer)73.6−3.558.1−15.9
The mechanism is backbone-dependent, and we say so. For Qwen3-VL-235B every ablated component is load-bearing, costing 10.4, 9.4 and 11.3 points. For Gemini-3-Flash the same interventions cost 0.6, 0.1 and 0.4 — all within a tie — so for the stronger backbone a simpler explanation survives: structured preprocessing with a capable model may account for most of the gain. We therefore restrict the strong per-stage claim to Qwen. Shuffling the temporal order of the same representation costs 11.8 points on Gemini and 19.9 on Qwen, which separates temporal order from unordered structure — but not object-centric structure from generic structure. The paper names the controls we lack.

Identity linking

Identity-linking and conflict-resolution ablation on the Qwen3-VL-235B pipeline.
ConfigurationAfaithAtarget
Base — Δtmax = 15 s, τconf = 0.75, bipartite54.274.0
Temporal constraint
Δtmax = 5 s (strict — fragments trajectories)50.471.2
Δtmax = 30 s (relaxed — false-positive matches)51.172.8
Confidence threshold
τconf = 0.90 (high precision)48.970.5
τconf = 0.50 (high recall)51.271.8
Conflict resolution
Greedy instead of bipartite matching47.569.4

Generality and cost

Beyond STRAND — and what it costs

External benchmarks (%). Since STRAND is built around object-centric trajectories, these results check generality rather than claim state of the art — and the picture is mixed.
MethodVideoHallucerVideo-MMEEgoSchema
Ours (Gemini-3-Flash)76.876.878.4
Ours (Qwen3-VL-235B)71.473.676.4
Gemini-3.1-Pro63.177.376.8
GPT-567.576.275.4
Qwen3-VL-235B59.872.672.1
LLaVA-OV48.566.360.1

The Gemini instantiation leads the strongest baseline by 9.3 points on VideoHallucer and 1.6 on EgoSchema, and trails Gemini-3.1-Pro on Video-MME by 0.5. The Qwen instantiation improves on its own extractor backbone, run end-to-end, by 11.6 points on VideoHallucer.

The gain is not free

Cost per video at Q = 39.7 questions. Trajectory construction is amortised: 12.2 extraction calls are paid once per video and reused across every query on it, then two calls per query, giving 91.6. A uniform 256 tokens per frame is assumed for all systems so the comparison reflects structural cost rather than a resolution choice. Rows marked * are API-served and their latencies include network overhead, so they are not commensurable with the locally served measurements.
SystemPre-proc. (s)Per-query (s) CallsInput tokensOutput tokens
Ours (Gemini-3-Flash)*14.23.291.6962,40018,100
Ours (Qwen3-VL-235B)18.42.991.6958,15717,450
Gemini-3.1-Pro*0.03.539.7654,4151,985
Qwen3.5-27B0.02.839.7654,4151,985
Cosmos-Reason2-8B0.02.139.7654,4151,985
VideoMind-7B6.54.240.7710,00014,000
TraveLER8.23.8119.1820,00012,000
SeViLA5.13.179.4690,0009,500

Our framework buys its 5.6-point target-accuracy gain over Qwen3.5-27B with more wall-clock time, more model calls and a larger token budget. That is the trade the budget-matched comparisons are designed to price.

Qualitative analysis

Right answer, wrong reason

The paired evaluation surfaces reasoning that target-only scoring would credit in full.

Two cases. Case 1: Gemini-3-Flash answers the target correctly but fails the identity sub-question, while our framework answers both correctly. Case 2: Gemini-3-Flash answers the sub-question correctly but gets the target outcome wrong, while our framework recovers the event order.
Two failure patterns. Case 1: a baseline answers the target correctly while failing the identity sub-question it depends on — a correct answer for the wrong reason. Case 2: the baseline reads the outcome of the game backwards; our framework recovers the order of events. In both, the trajectory memory records the state changes the target turns on.

These cases illustrate why target accuracy alone can overstate faithful video understanding. The quantitative form of the claim is the gap between Atarget and Afaith in the main table: Gemini-3.1-Pro answers 61.4% of targets correctly but supports only 38.9% of them fully.

Benchmark items

Four examples

Each target is paired with the atomic sub-questions it rests on. Under Faithful Accuracy, every one of them has to be right. Play a clip and the dependency becomes plain: the answer turns on something that exists only across time.

Pure temporal ordering

Which came first?

Both events happen, so recognising them is not enough — only their order decides the answer.

Are the cookies smashed before the strawberries?No

Supporting sub-questions

  • Are the strawberries smashed in the video?Yes
  • Are the cookies smashed in the video?Yes
  • Are the cookies smashed after the strawberries?Yes

Identity through occlusion

Which object ended up where?

Containers are lifted, covered and swapped, so each object’s identity has to survive the occlusion rather than be read off the final frame.

Is the white object under the brown object after the woman lifts up both objects?Yes

Supporting sub-questions

  • Did the woman lift up the white object?Yes
  • Did the woman lift up the brown object?Yes
  • Is the white object below the brown object after lifting?Yes

Paired target

Is the yellow object under the brown object after the woman lifts up both objects?No

State tracking across repeated trials

Which door, in which run?

The run is repeated with a different configuration each time, so the outcome of the first experiment cannot be recovered from a later state of the apparatus.

In the first experiment, did the yellow ball go to door number 2?Yes

Supporting sub-questions

  • Was there a yellow ball in the first experiment?Yes
  • In the first experiment, did the yellow ball reach door number 2?Yes

Paired target

In the first experiment, did the yellow ball go to door number 1?No

Actor disambiguation over time

Who went first?

A two-team game. The paired target swaps a team colour for a player description, so answering it means linking an individual to a team and to the order of play.

Does the blue team start playing first?Yes

Supporting sub-questions

  • Is there a blue team in the video?Yes
  • Did the blue team begin playing before any other team?Yes

Paired target

Does the black-haired woman start playing first?No