embodied-vision-language-models-trained-separate-capability-specialists-spatial-temporal CLM-0564For embodied vision-language models trained as separate capability specialists (spatial, temporal, action guidance, state verification) from one backbone, consolidating them by TIES weight merging and then routed multi-teacher on-policy distillation retains more of each specialist's gain than either step alone or than mixed-data RL, which shows capability-specific biases.
Capability: Turning plans into physical action
Observed on
Two Qwen backbones (35B-A3B MoE, 2B dense); specialists trained with GRPO on verifiable rewards; retention measured on eight signature benchmarks under matched prompts and scorers..
Sources
- Validated by: Controlled ablation on 35B-A3B against the shared backbone, four standalone experts, Mix-RL, TIES-only and MOPD-only across eight capability benchmarks; TIES+MOPD best on 3 of 8 and second on the other 5, improving 7 of 8 over MOPD alone. Also 28/34 and 30/34 matched benchmark rows improved over the Qwen initialisations, and EB-HAB success 46.0 to 63.0.. Measured on the authors' own harness with matched decoding per pair; [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When adapting a pretrained VLA manipulation policy with only a handful of target demonstrations, retrieving prior demonstrations at the level of LLM-decomposed subtasks (then reranking by behavior-feature similarity) raises success more than retrieval by visual, optical-flow, or state-action similarity, and the margin is concentrated on long-horizon composite tasks rather than single-skill pick-and-place.Turning plans into physical action · reviewed by AI
- In multi-agent literature-synthesis pipelines, most of the gain over a fixed role-based pipeline comes from the shared-workspace architecture and the added critique/dissemination actions rather than from learning the coordination policy: on OARelatedWork, a learned decentralized RL controller scored 3.82 overall versus 3.80 for a centralized LLM coordinator given the same workspace and action space, while a naive orchestrator without that architecture scored 3.40.Getting more from several models than from one · reviewed by AI
- When open-source vision-language models are post-trained with RL or critique tuning to act as judges of chart-to-code outputs, they become more likely to output Accept rather than better at verification: pairwise-comparison accuracy barely improves under an order-swapped protocol, while acceptance rates rise and specificity falls, so they ratify flawed but plausible-looking charts.Judging another's work as well as a careful reviewer would · reviewed by AI
- When RL-training a central-agent multi-agent harness with 1.5B-3B Qwen2.5 backbones on multi-hop QA, optimizing only the coordinating central policy and leaving sub-agents frozen scores higher than jointly training central and sub-agent policies (average F1 47.79 vs 45.36, accuracy 42.88 vs 40.49 at 3B).Getting more from several models than from one · reviewed by AI
- In embodied multi-agent household planning with a fixed high-level skill interface, a team's task success rate overstates coordination quality, because goal predicates can be reached through illegal shortcuts: several of 11 tested MLLMs succeed on more episodes than they produce legal plans on (Claude Haiku 4.5 63.1% success vs 47.2% legal; Qwen3.6-Plus 69.1% vs 56.7%), while others honour constraints yet miss the goal.Getting more from several models than from one · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A replication where TIES-only or MOPD-only, or plain mixed-data RL, retains specialist gains as well as or better than TIES+MOPD across all four capability families, or where the ordering does not hold on a different backbone.