gtr-turbo-vlm

name	gtr-turbo-vlm
title	GTR-Turbo: Merged Checkpoint as Free Teacher for Agentic VLM Training
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.13043
keywords	["vision-language","reinforcement-learning","teacher-free","distillation","checkpoint-merging"]
description	Eliminate expensive external teacher dependencies in VLM RL training via merged-checkpoint teachers. Uses TIES merging of historical RL checkpoints to create free, stable teacher models for step-level guidance—matching external teacher performance while reducing training time 50% and computational costs 60%.

Overview

GTR-Turbo shows that effective VLM RL training doesn't require external teachers—historical checkpoint merging provides guidance automatically.

Core Technique

Checkpoint Merging via TIES:

class CheckpointMergingTeacher:
    def __init__(self, checkpoint_history):
        self.checkpoints = checkpoint_history
        self.merged_model = None

    def merge_checkpoints(self):
        """
        TIES merging: Trim, Elect Sign, Merge for stable teacher.
        """
        # Collect weight differences from reference
        weight_diffs = []
        for ckpt in self.checkpoints:
            diff = ckpt.weights - reference_model.weights
            weight_diffs.append(diff)

        # TIES: Element-wise analysis
        merged = reference_model.weights.clone()
        for param_name in merged.state_dict():
            # Collect this parameter across checkpoints
            param_diffs = [wd[param_name] for wd in weight_diffs]

            # Vote by sign (elect sign)
            signs = torch.sign(torch.stack(param_diffs))
            elected_sign = torch.mode(signs, dim=0).values

            # Average magnitude
            magnitudes = torch.abs(torch.stack(param_diffs))
            avg_magnitude = torch.mean(magnitudes, dim=0)

            # Merge: elected sign * average magnitude
            merged[param_name] = elected_sign * avg_magnitude

        self.merged_model = create_model_from_weights(merged)
        return self.merged_model