LoRA adapter on Qwen/Qwen3-VL-4B-Instruct. This is the on-policy DPO stage of BirdAgent (see the flagship Chinzhu/BirdAgent-Qwen3VL-4B GSPO model for the full description, tools, results, and figures).

📄 Paper (under review) · 💻 Code · 📊 Benchmarks

Stage: on-policy DPO on top of SFT, optimizing soft preferences only — ledger discipline, call parsimony, turn ordering — never grain or abstention (those belong to RL). Pairs are perturbed SFT gold trajectories (chosen = gold; rejected = gold + one redundant call, verdict unchanged).
What it buys: calibration sharpening — species-declaration precision 0.33 → 1.0 (the agent stops naming species it cannot support). Pooled solve 0.34 overall (beats Sonnet, second only to Opus among all conditions).
Adapter: LoRA r=64, alpha=128, dropout=0.05 (trained bf16, adapter saved fp32); targets language-model q/k/v/o/gate/up/down_proj.
Note: this checkpoint is the initialization for GSPO; the released agent is the GSPO model, which adds orchestration on top (pooled 0.34 → 0.35). Same intended use, limitations, and responsible-use terms as the flagship card.

Citation

bibtex

@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}
}

Chinzhu/BirdAgent-Qwen3VL-4B-DPO

BirdAgent — Qwen3-VL-4B agentic bird identifier (on-policy DPO)

Citation

Chinzhu/
BirdAgent-Qwen3VL-4B-DPO