Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-scoring-128k-balanced-pl-naive

Qwen3-VL-8B-Instruct-UI-Genie-scoring-128k-balanced-pl-naive

A scalar reward model for mobile GUI agent steps. Plackett-Luce fine-tune of a UI-Genie-trained Bradley-Terry stage-1 base, using Latent Peer Voting (LPV) labels built from the *_all_naive construction.

What it is

• —Base: qwen3vl_8b_128k_balanced_bt, a 128k-balanced UI-Genie BT reward model over Qwen/Qwen3-VL-8B-Instruct.
• —Objective: Plackett-Luce on LPV labels, all_naive label construction, expert weights average.
• —Checkpoint: epoch 4 of 8, LoRA rank 64 over all linear layers with the scalar score head in modules_to_save; merged here.
• —Input contract: ui-genie-rm-paired-v1 - a mobile_use tool-spec system turn carrying the logical screen size, the executed history with the recent screenshots interleaved, and the candidate action as a closed <tool_call>. Training and evaluation used this same contract, 16,384 tokens, no truncation.

Usage

Score is score_head(last_token_hidden_state); higher is better. Coordinates must be in the logical-screen pixel space declared by the system prompt.

Evaluation

AndroidFlux recovery (76 fixed pairs, history screenshots attached) and the UI-Genie 1,000-pair set. clean 52 excludes 7 pairs whose two candidates land within 25px of each other and 17 whose source action is outside the pinned tool spec; error_path carries step-level labels, clean_path carries episode-level credit projected onto one step.

Read these with the caveats. No recovery difference here is statistically significant: the clean 52-pair subset has a standard error of about 3.6 pairs, and 24 of its 52 pairs are decided by a near-constant preference for terminate. The UI-Genie change is significant (n=1000, standard error about 9 pairs) and it is a loss - LPV fine-tuning on AndroidFlux costs UI-Genie accuracy, more so for the naive label construction than any other setting tested.

The score head moved very little during fine-tuning (max absolute change 0.001, cosine similarity 0.9998 against the base head); the behaviour change comes from the merged LoRA.