TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

TinyAstral evolves a teacher’s harness, then distills it into a 4B agent that matches the teacher’s success on RoboTwin using roughly one fifth the accelerator time.

https://t.co/lhKx4d5BB8 https://t.co/MFXFXaOPrm

TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

Source: https://arxivb.org/abs/2610.41216 [Submitted on30 Oct 2026]

Selection snapshot savedNov 1, 22:45 PT. Title, authors, abstract and submission date are frozen here.

Kehong Gong, Xiaoxu Zheng, Guanli Hou, Seunghun Lee, Mengqi Guo, Runjia Tan, Zhengyu Wen, Qi Wang, Xiaoyu He, Mingyuan Zhang, Michael Bi Mi

Awarded11/01/2026 11:14pm PST

83/100Editorial score

TinyAstral evolves a teacher’s harness, then distills it into a 4B agent that matches the teacher’s success on RoboTwin using roughly one fifth the accelerator time.

Waiting for 24-hour stats

Abstract

Robot agents built on vision-language models depend on both the planner and its harness: the tools, feedback, memory, and recovery logic around it. Can improvements to a large teacher’s harness be carried into a compact agent with lower inference compute cost? We present TinyAstral, a two-stage approach. We first freeze an open-weight teacher and improve its harness through recursive self-improvement (RSI): frontier LLM agents read the teacher’s trajectories and propose harness edits, and each edit is accepted only if it passes a paired statistical test. We then freeze the final harness, collect teacher trajectories with reasoning and failure recovery, and distill a small student that replaces the teacher in the same harness. On RoboTwin, harness evolution raises the teacher’s success from 33.4% to 63.6%; the final 4B student reaches 62.8%. Across four harness versions, distilled 4B students score close to their corresponding teachers, and later versions yield stronger students, showing that harness evolution gains carry into compact models. The harness gains also transfer to other frozen planners without retraining. On identical episodes, the final 4B student outperforms open 4B agents, the open 27B harness-RSI model SenseNova-RoboRSI, and our own teacher model placed in three competitors’ harnesses. It stays close to its teacher on all three training benchmarks, and on three benchmarks never used for RSI or training it scores close to a 552B planner running the same harness. On the measured deployment episodes, it uses about one fifth of its teacher’s accelerator time.

Editorial assessment

83/ 100

Automated assessment of the title, authors and abstract. It can be wrong.This assessment was saved with the selection.

Core ideaThe paper decouples agent optimization into two steps: recursively evolving the agent’s harness (tools, context, recovery logic) using a larger teacher with acceptance tests, and then distilling the teacher’s trajectories within that optimized harness into a compact 4B parameter student.Why it mattersBuilders wrestling with inference latency and cost in agentic workflows get a clear architectural blueprint: instead of fine-tuning large models directly, evolve the agentic scaffolding first and distill the resulting high-quality trajectories into edge-sized models.The ranker’s takeThe core insight here is brilliant: treat the agent scaffold itself as an evolving software artifact before distilling the policy down to a lightweight model. Most distillation pipelines freeze the prompt or tools and just compress model weights, but iterating on harness logic first appears to provide massive leverage. I am particularly impressed by the 4B student retaining nearly all the gains of the evolved harness while cutting compute by roughly 80%.ReservationThe abstract does not detail the nature of the frontier LLM doing the harness edits or how sensitive the statistical acceptance test is to simulation vs. physical domain shift.What to check nextInspect what kinds of harness edits were actually discovered (e.g., prompt tweaks vs. complex tool orchestration) and whether the evaluation spans real-world robots or strictly simulated benchmarks like RoboTwin.Discussion hookDoes optimizing the software scaffold around an agent yield higher performance ceilings than scaling the base model’s parameter count for specific agentic loops?Assessment confidence:medium· AssessedNov 1, 19:12 PT

Ranker:Gemini editorial judgment· v2.0.0Authorsarxivbangers contributorsMethodeditorial-v2Fingerprintead79b0bec93c81d487bfa06d5188d7edbd2e5faa3fb2c4506e2d88a5d36002b

Discussion