[Ryenhails]

This PR adds results of VisAgentRet (EVIE-4.5B + GLM-5.3-Flash) to the ViDoRe V3 pipeline benchmark.

GLM-5.3-Flash runs as an agent in the Codex CLI. It searches EVIE-4.5B with its own queries and looks at page images. Then it returns its top 10 pages. No OCR text is used.

The average nDCG@10 over the 8 English datasets is 0.7001. Search takes 24.4 s per query with 16 concurrent queries on one B300.

The CLI writes num_queries only per language, so we copied the English value into overall for the CI check. The pipeline implementation is available at visagentret.

[QuentinJGMace]

Hey @Ryenhails, thank for the submission it looks great ! Just a few minor comments:

• You ran evals with num_concurrent=16. Other entries ran sequentially, so this isn't comparable on the leaderboard. Could you run a short sequential timing (e.g. 30–50 queries)? Otherwise I'll add a footnote.

• The shell tool is still available to the agent (disabled by the prompt only). Could you report the total n_commands over all episodes? (should be 0 I guess)

• The agent sees doc_id, which is often a descriptive title in addition to the actual doc, could you mention it in the description file ?

• Can you point to nvidia/GLM-5.3-Flash-NVFP4 rather than zai-org/GLM-5.3-Flash which seems to be the model you used and add an info_url to your README for clarity ?

I'll fix the CLI for the num_queries patch

[Ryenhails]

Thanks for the review!

You ran evals with num_concurrent=16. Other entries ran sequentially, so this isn't comparable on the leaderboard. Could you run a short sequential timing?

We ran the first 30 queries of hr with num_concurrent=1 on the same B300, with the same code as the submitted run. Search takes 49.3 s per query. We added this to the hardware notes.

The shell tool is still available to the agent (disabled by the prompt only). Could you report the total n_commands over all episodes?

The total n_commands over all 2419 episodes is 0.

The agent sees doc_id, which is often a descriptive title in addition to the actual doc, could you mention it in the description file?

Done. The description now says that the search results show the doc_id of each page.

Can you point to nvidia/GLM-5.3-Flash-NVFP4 rather than zai-org/GLM-5.3-Flash which seems to be the model you used and add an info_url to your README for clarity?

Done. The agent node now points to nvidia/GLM-5.3-Flash-NVFP4, and info_url points to the README. The pipeline_args in the result files keep zai-org/GLM-5.3-Flash, because that is the served model name we gave vLLM.

I'll fix the CLI for the num_queries patch

Thank you!

[QuentinJGMace]

Looks good ! merging

[Copilot]

🟢 Approval recommended

The data-only additions have consistent submission structure, query counts, timing calculations, and reported averages, with no blocking issues identified.

No open issues 🎉

What changed in this PR

Adds VisAgentRet (EVIE-4.5B + GLM-5.3-Flash) to the ViDoRe V3 pipeline benchmark.

Changes:

• Adds results for English queries across eight datasets.
• Documents pipeline components, hardware, concurrency, and timing methodology.
• Includes overall query counts for CI validation.

File	Description
results/​pipeline_descriptions/​visagentret_evie-4.5b_glm-5.3-flash/​description.json	Pipeline architecture and execution details.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_physics.json	Physics metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_pharmaceuticals.json	Pharmaceuticals metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_industrial.json	Industrial metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_hr.json	HR metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_finance_fr.json	French finance corpus metrics for English queries.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_finance_en.json	English finance corpus metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_energy.json	Energy metrics and timings.
results/​metrics/​visagentret_evie-4.5b_glm-5.3-flash/​vidore_v3_computer_science.json	Computer science metrics and timings.

_{Review effort: Balanced}