Entity · paper

When Model Merging Rivals Joint Multi-Task Reinforcement Learning: A Task-Vector Geometry Analysis

paperactiveprovisionalwhen-model-merging-rivals-joint-multi-task-reinforcement-learning-a-task-vector-geometry-analysis-fee72dc0·1 events·first seen 4h ago

Aliases: When Model Merging Rivals Joint Multi-Task Reinforcement Learning: A Task-Vector Geometry Analysis

Co-occurring entities

RAM+TIES LOOP Qwen3-4B AppWorld

More like this (12)

Steerable Model Merging Multi-Task Learning Model Merging / Weight Interpolation Geometric Action Model for Robot Policy Learning Geometric Trajectory and Contrastive Learning When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It Physics-EnhAnced Reinforcement Learning Geometric Action Model Reward Modeling for Multi-Agent Orchestration Where Does the Answer Come From? Benchmarking View-Level Visual Evidence Identification in Multi-View MLLMs for Autonomous Driving Task and Motion Planning

Recent events (1)

6arXiv · cs.AI·4h ago·source ↗

Model merging matches joint multi-task RL training on AppWorld benchmark, explained by near-orthogonal task vectors

A new arXiv paper provides the first direct comparison of model merging versus joint multi-task reinforcement learning training, using Qwen3-8B specialists trained on the AppWorld agent benchmark with the LOOP algorithm. Merging methods (TIES, RAM+) statistically match jointly trained models on task-goal completion. The authors explain this via task vector geometry: specialist task vectors are near-orthogonal (cosine similarity 0.06–0.10) despite ~65% parameter support overlap, causing sign- and support-based merging methods to collapse to near-uniform averaging.

Evaluation and Benchmarking Agent and Tool Ecosystem RAM+When Model Merging Rivals Joint Multi-Task Reinforcement Learning: A Task-Vector Geometry Analysis TIES +4 more