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

Source: [arXiv](https://arxiv.org/abs/2608.06361v1)  
Feed7 permalink: https://feed7.dev/p/2608-06361v1-1n3dr85  
Published: 2026-08-06T17:57:06.000Z  
Trust: Needs Review (needs_review)

## Why Included

Controlled traces show video models can improve final counting scores without faithfully recovering events, so agents handling video need timestamp-level checks, not answer-only evals.

## Source Summary

The study profiles event counting across **2,190 controlled videos** with executable traces. Gemini 3.6 Flash reached 80% reliability for persistent transitions up to **12 events**, but had no reliable positive-count region for transient blinks.

## Practical Implication

For video agents, evaluate the reported event sequence against timestamps, not just the final count. In the high-count, high-frequency regime, only **0.2%** of counts were correct and **18.1%** of true events were recovered.

## Agent-Ready Context

The study profiles event counting across **2,190 controlled videos** with executable traces. Gemini 3.6 Flash reached 80% reliability for persistent transitions up to **12 events**, but had no reliable positive-count region for transient blinks.

For video agents, evaluate the reported event sequence against timestamps, not just the final count. In the high-count, high-frequency regime, only **0.2%** of counts were correct and **18.1%** of true events were recovered.

More frames raised Bounce Ball accuracy from 19.6% to 29.3%, while full sequence agreement remained 3.7%. Prompt changes also brought limited gains, so higher aggregate accuracy may conceal poor event bookkeeping.

## Connected Context

Feed7 judgment across 390 accumulated Signals:

This makes temporal bookkeeping a distinct video-agent capability that aggregate counting accuracy can conceal. It reinforces the candidates’ case for inspecting trajectories and partial progress, but supplies a stronger executable oracle: compare every reported event with timestamped ground truth. More frames and prompt changes offer limited relief, narrowing the value of simple inference-time adjustments for transient or dense events.

- [Rethinking Environments for Long-Horizon Work — Rayan Garg, Theta Software](https://feed7.dev/p/rethinking-environments-for-long-horizon-work-rayan-garg-theta-software-11r7wbx) — The timestamped event sequence is a concrete form of the queryable trajectory this candidate calls for, and it reveals failures hidden by final-count grading.
- [SocietyBench: Forecasting Counterfactual Social-World Evolution](https://feed7.dev/p/2608-04009v1-05m5u8w) — Both separate temporal performance from a broader aggregate result, reinforcing event-level reporting when averages can hide sharply different reliability regimes.
- [Teaching AI to Find Real Vulnerabilities — David Brumley, Bugcrowd](https://feed7.dev/p/teaching-ai-to-find-real-vulnerabilities-david-brumley-bugcrowd-1ok0f7q) — Executable traces play the role of a deterministic oracle, while event recovery measures partial progress much as a capability ladder does when the final answer is wrong.

## Context Map

- Layer: benchmark
- Domains: video
- Topics: agent-evals, benchmark-integrity, agent-reliability

## Uncertainty

- More frames raised Bounce Ball accuracy from 19.6% to 29.3%, while full sequence agreement remained 3.7%. Prompt changes also brought limited gains, so higher aggregate accuracy may conceal poor event bookkeeping.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.