Skip to content

LVBench pilot: 100 questions over 17 videos

Generated 2026-09-12 by eval/report.py from 3 run file(s). Accuracy is exact match on the option letter; intervals are 95% Wilson. Costs are provider list prices per question; indexing cost is not included for the index-based configurations.

Configuration n accuracy 95% CI unparsed cost / q tokens / q tool calls / q latency p50 latency p95 no-decode citation in range
agent 100 68.0% 58.3–76.3 1 $0.097 28,216 4.32 14.8 s 33.8 s 37% 64% (n=98)
retrieval-only 100 50.0% 40.4–59.6 1 $0.025 6,763 1.00 7.3 s 13.3 s 100% 54% (n=94)
uniform-32 (claude-sonnet-5) 100 52.0% 42.3–61.5 1 $0.038 12,294 0.00 10.9 s 16.6 s 100%

Reading the pilot

  • Scope. 100 questions stratified by task type over the 17 LVBench videos that had been downloaded and indexed at the time (YouTube's bot check paused the rest); the 95% intervals are wide and the per-type cells small. Treat this as a pipeline check and a first signal, not the LVBench number.
  • The agent leads by 16 to 18 points over both baselines at about 2.5 to 4× their cost per question. Its gains are largest on temporal grounding (92% vs 42%) and reasoning (92% vs 54%): questions where finding the moment matters. On entity recognition it is ahead of both, and summarization (n=5) is weak for every configuration.
  • Retrieval-only equals the uniform-frames baseline at half the cost: an index and one search buy what 32 uniformly sampled frames plus the caption file buy, without decoding anything at question time.
  • Two harness fixes came out of the first pass of this pilot: the agent returned empty text after tool-heavy turns and the retrieval-only policy wrote pseudo tool calls when tools were withheld. Final turns now keep the tool definitions with tool_choice: none (accuracy 58% to 68% and 18% to 50% on the same questions).
  • Citations. 64% of the agent's answers carry a citation within 60 s of LVBench's time_reference; the benchmark does not ask for one, so this is a bonus signal of grounding.
  • Cost. Index-based rows exclude indexing: 17 videos (about 19 hours) took 45 minutes of machine time on the GPU build and no API calls.

Accuracy by task type

Task type agent retrieval-only uniform-32 (claude-sonnet-5)
entity recognition 64% (n=47) 49% (n=47) 43% (n=47)
event understanding 67% (n=45) 53% (n=45) 53% (n=45)
key information retrieval 75% (n=12) 67% (n=12) 58% (n=12)
reasoning 92% (n=13) 54% (n=13) 77% (n=13)
summarization 40% (n=5) 0% (n=5) 40% (n=5)
temporal grounding 92% (n=12) 42% (n=12) 42% (n=12)

Tool-call profiles

agent: search > search > view ×16; search > search ×5; search > view ×4; search > search > search > search > timeline > search > search ×4; search > search > search > search > search > search ×3; search > search > view > view > describe > view ×3; search > search > search > search > search > view ×3; search ×3

retrieval-only: search ×100

uniform-32 (claude-sonnet-5): (none) ×100

Runs

  • agent: /data/videoindex/eval/runs/lvbench-pilot-agent.json — {"benchmark": "lvbench", "policy": "agent", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 100, "seed": 1, "vi_version": "vi 0.1.0", "started": "2026-09-12T07:43:18Z"}
  • retrieval-only: /data/videoindex/eval/runs/lvbench-pilot-retrieval.json — {"benchmark": "lvbench", "policy": "retrieval-only", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 100, "seed": 1, "vi_version": "vi 0.1.0", "started": "2026-09-12T07:50:36Z"}
  • uniform-32 (claude-sonnet-5): /data/videoindex/eval/runs/lvbench-pilot-uniform32.json — {"benchmark": "lvbench", "policy": "uniform-32", "model": "claude-sonnet-5", "frames": 32, "transcript": true, "sample": 100, "seed": 1, "started": "2026-09-12T07:54:07Z"}

Questions per run: up to 100. Benchmark videos are public YouTube content and may be in model training data; read per-configuration deltas rather than absolute scores.