The Low-Frequency Trap: Video–Language Models Fail at Simple Event Bookkeeping

Trace-Grounded Parametric Profiling under Varying Event Count and Frequency

1University of Maryland, College Park 2Scale AI

* Equal contribution

Figure 1: Controlled Trace-Grounded Paradigm and Low-Frequency Trap

Figure 1: Controlled trace-grounded paradigm and the Low-Frequency Trap. Real-world benchmarks entangle visual complexity with event dynamics, hiding failure causes. Our evaluation isolates event count (N) and event frequency (F) across three controlled tasks: bouncing balls (transient wall contacts), blinking lights (transient visual flashes), and state machine indicators (persistent state changes). Paired with executable ground-truth traces, our benchmark evaluates both the final count and event-by-event temporal execution.

Abstract

Real-world video benchmarks provide broad coverage, but their fixed clips entangle event count, rate, duration, and visual complexity, making failure modes hard to isolate. While existing programmatic benchmarks offer better control, they primarily score only the final answer rather than auditing reported events against executable ground truth.

To bridge this gap, we introduce trace-grounded parametric profiling for event counting in three controlled video tasks: bouncing-ball wall contacts, visual blinks, and categorical state transitions. Across 2,190 videos, we systematically vary event count (N) and frequency (F) while holding rendering fixed. Each video includes an executable event trace for capability-surface estimation and timestamp-level evaluation.

Our results reveal a staged temporal failure: at an 80% reliability threshold, Gemini 3.6 Flash reliably counts persistent state transitions up to 12 events at 0.5 and 1.0 Hz, yet demonstrates no reliable positive-count region for transient blinking events. Thus, event representation (persistent vs. transient) dictates whether a model initially accesses evidence—a limitation that compounds as count and frequency increase. In high-count, high-frequency regimes, only 0.2% of final counts are correct and models recover just 18.1% of true events. Furthermore, increasing sampling rate boosts Bounce Ball final accuracy from 19.6% to 29.3%, yet reported sequences agree with ground truth only 3.7% of the time, proving that extra frames inflate final scores without producing faithful event recovery.

2,190
Evaluated Videos
Systematic grid sweep across N × F parameter space.
0.2%
High-Load Accuracy
Final count accuracy under high event count and frequency.
18.1%
True Event Recall
Model ability to recover true timestamped events under load.
3.7%
Trace Agreement
Timestamp sequence alignment despite 4× frame density boost.

Controlled Profiling & Trace Evaluation

Figure 2: Benchmark Task Overview

Figure 2: Benchmark tasks and trace-grounded evaluation. Controlled synthetic videos generated across three task domains: Bounce Ball (wall contact bookkeeping), Blinking (transient visual flashes), and State Machine (persistent state transitions). Each sample includes exact event timestamps, enabling evaluation of both final-answer correctness and step-by-step trace fidelity.

Empirical Capability Surfaces

Gemini 3.6 Flash
Figure 3a: Gemini 3.6 Flash Capability Surfaces
Qwen 3 VL 235B Instruct
Figure 3b: Qwen 3 VL 235B Instruct Capability Surfaces

Figure 3: Capability surface heatmaps across N × F space for frontier VLMs. Reliability boundaries for Gemini 3.6 Flash (top) and Qwen 3 VL 235B Instruct (bottom) across event count (N = 1..12) and frequency (F = 0.1..2.0 Hz). Persistent state transitions remain reliable (accuracy ≥ 80%) up to high event counts, whereas transient blinking events demonstrate early failure boundaries even at low event loads.

Diagnostic Interventions

Figure 4: Diagnostic Interventions Analysis

Figure 4: Frame density, keyframes, and prompt interventions. Increasing sampling rate (e.g. from 1 fps to 4 fps) improves final-answer accuracy on Bounce Ball from 19.6% to 29.3%. However, auditing reported timestamp traces reveals that trace agreement remains at a dismal 3.7%. Extra frames allow models to guess final counts better without executing faithful temporal bookkeeping.

Real-World Transfer

Figure 5: Real-World Transfer Evaluation

Figure 5: Transfer evaluation on natural videos. Testing models on natural repeated-event videos confirms that performance degradation concentrates at high event counts, validating that controlled synthetic findings reflect real-world video-language capabilities.

Qualitative Examples

📹

Faithful Event Recovery (Correct Match)

Bounce Ball (N=5, F=1.0Hz)
Example 1 of 18
1.00x
QUESTION PROMPT

GROUND TRUTH EXECUTABLE TRACE GT COUNT: 5
Executable Timestamps:

        
GEMINI 3.6 FLASH PREDICTION
Trace F₁: 100.0% Precision: 100.0% Recall: 100.0% PRED ANSWER: 5
Raw Model Output:

          
DIAGNOSTIC JUSTIFICATION

BibTeX Citation

If you find our work useful in your research, please cite:

@misc{baskar2026lowfrequencytrapvideo,
      title={The Low Frequency Trap: Video Language Models Fail at Simple Event Bookkeeping}, 
      author={Sarvesh Baskar and Zikui Cai and Shayan Shabihi and Anirudh Satheesh and Muhammad R. Islam and Udari Madhushani Sehwag and Tom Goldstein and Furong Huang},
      year={2026},
      eprint={2608.06361},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.06361}, 
}