Chinzhu/BirdAgent-Benchmarks

BirdAgent benchmarks — agentic & calibration

📄 Paper (under review) · 💻 Code · 🧠 Model Chinzhu/BirdAgent-Qwen3VL-4B · 🏋️ Training data Chinzhu/BirdAgent-Train

These are the held-out evaluation sets. For the training data (SFT / DPO / GRPO trajectories) see Chinzhu/BirdAgent-Train.

Two open, tier-stratified benchmarks for agentic, tool-orchestrating bird identification. Unlike monolithic recognition sets, these isolate orchestration: every item is built so that a strong single-image model must lose, and the deciding evidence is reachable only by calling the right tool.

• agentic — 500 items, each with an engineered information gap (scenarios S1–S4 below) that a tool-free model cannot close. Every item's minimal necessary tool set (mnts, 1–2 tools) is hand-verified (mnts_confirmed = true). v0 items are single-gap (one deciding tool); multi-gap and fault-injection layers are left to a future release.
• calibration — 400 items, testing when to commit to a species vs. degrade to genus/family or abstain. Each item carries a reachability label — whether the true species is actually within a tool's top-k — which makes the tool ceiling explicit: 303/400 (76%) are truth_missing (no tool can surface the truth), the rest top1_high / mid / top1_low.

Both sets are stratified by frequency tier (common / uncommon / rare): agentic ≈ 181 / 178 / 141, calibration ≈ 134 / 133 / 133.

The two splits use different schemas (the agentic set is scenario-based; the calibration set is reachability-based) — see the field tables below.

Fields

Common to both splits: id · modality (image / sound / image_sound) · query · mnts (minimal necessary tool set) · freq_tier (common/uncommon/rare) · geo_band · true_species_ebird (eBird code — scoring only, not shown to the model) · image (iNaturalist URL, null for sound-only) · audio (Xeno-canto URL, null for image-only).

agentic only:

field	meaning
scenario	S1 look-alike (voice-diagnostic) · S2 range/season prior · S3 poor image quality · S4 multi-bird / clutter
difficulty	L1 (single-gap; the only layer in v0)
mnts_confirmed	MNTS hand-verified (always true here)
privileged_geo	true {lat, lon, date} — held out of the prompt; the model must call geo_prior to use location
cf_geo	a counterfactual {lat, lon} for the anti-memorization probe (a model that truly uses the tool changes its verdict)

calibration only:

field	meaning
purpose	"calibration"
reachability	is the truth reachable through a tool? top1_high · mid · top1_low · truth_missing
lat, lon, date	observation geo/date (scoring / tool input; keep out of the prompt)

Scenarios (agentic set)

scenario	information gap	required orchestration
S1 look-alike, voice-diagnostic	near-identical plumage (e.g. Empidonax)	classify_sound on the diagnostic call
S2 range/season prior	candidate looks right but is implausible here/now	geo_prior(lat,lon,date) to down-weight
S3 poor image quality	distant / blurry / backlit; field marks hidden	quality_gate → zoom/enhance → reclassify
S4 multi-bird / clutter	several birds or a tiny subject	detect_bird → crop → classify the isolate

Each scenario has 125 agentic items. Abstention / out-of-distribution behavior is tested by the separate calibration set (via reachability and the option to degrade or abstain), not as an agentic scenario.

Protocol (important)

• Privileged information is never shown. GPS, date, and the true taxon are used only to select and score. A model that wants a range prior must call the geo_prior tool — this is what separates orchestrates from was fed the answer.
• Identical-harness comparison. Score every model through one identical tool harness (same signatures, return schemas, call budget). Conditions: C0 bare · N vendor web search · C1 our exact tools, no orchestration training · C2 a trained agent · T0 image→classify_image (recognition floor).
• Primary metric: solve — correct at the declared grain (a genus verdict counts iff the genus is right; over-committed species are penalized). Also report species-declaration precision, overclaim, and mean calls.
• De-leaking. Items favor post-cutoff observations, perceptual-hash de-duplicated, media-ID de-leaked against the recognizers' pre-training pools.

Media

Raw media bytes are not redistributed. Each item carries its source image (iNaturalist open-data) and/or audio (Xeno-canto) URL; fetch them yourself, e.g.:

import json, requests
for line in open("agentic.jsonl"):
    it = json.loads(line)
    if it["image"]:
        img = requests.get(it["image"]).content   # iNaturalist photo
    if it["audio"]:
        wav = requests.get(it["audio"]).content    # Xeno-canto recording

Media licenses follow their sources (citizen-science; many are non-commercial), hence the cc-by-nc-4.0 label for this annotation package. Taxonomy is on the eBird/Clements crosswalk.

Citation

@inproceedings{wang2026birdagent,
  title     = {BirdAgent: A Small Vision--Language Model that Orchestrates
               Domain Tools Beats Large Models that Merely Hold Them},
  author    = {Wang, Xinzhu},
  booktitle = {Under review},
  year      = {2026}
}