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.
Figures & tables
Figure 2: Overview of TinyAstral. Stage 1: humans specify the rules and budget in advance. A master of frontier LLM agents reads trajectories of the frozen teacher acting through the harness (tools, feedback, memory, recovery) and proposes harness edits, which are accepted only if a paired A/B test on selection seeds passes a pre-registered statistical gate; model weights never change. Stage 2: after evolution ends, we freeze H∗ and collect teacher trajectories, including reasoning and recovery branches. These are mixed with offline real-robot supervision to train tiny students (0.8B–4B), which replace the teacher in the same harness. Evaluation covers three training domains and three unseen benchmarks, plus three closed-loop real-robot tasks on an SO-ARM101 arm.
Figure 3: TinyAstral-4B recovers on its own (original evaluation runs, final harness H∗ ; blue labels: zero-shot). Top two rows: closed loop on a real SO-ARM101 arm; middle two: training domains; bottom two: zero-shot. Cards: tool calls and returned codes verbatim from the logs; amber: the decisive step; bubbles: the student’s reasoning before it. More examples, including Guava and a distilled-vs-untrained comparison on one episode, are in Figs. 4 and 5 .
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Tool calls
Basis
RoboTwin
30
set by us
Guava
25 / 30
official per-task limits [ 27 ]
LIBERO-PRO
40
set by us
SimplerEnv
25
set by us, as Guava
VLABench
40
set by us, as LIBERO-PRO
EB-Manipulation
25
set by us, as Guava
Appendix
Table 6: Per-episode tool-call limits of our harness.
closed model; frozen RoboTwin-SFT VLA executor; success flag
60.0
Appendix
Table 7: Published results on our six benchmarks (success %), each reported by the cited work under its own protocol and episodes; the TinyAstral-4B value of Tab. 2 is given in each block header. Condition columns as in Tab. 2 ( × : advantage used by the system); Train data: — none, T teacher rollouts, E expert demonstrations, O other datasets. l Official RPent leaderboard ( https://rpent.readthedocs.io/en/latest/rst_source/leaderboard/performance.html , accessed 2026-10-23). v As collected in StarVLA [ 39 ] .
Figure 4: More recoveries of TinyAstral-4B (conventions as in Fig. 3 ). (1) RoboTwin: the card box is found only by a visual description; the basket handle slips with either arm, so the student grasps the rim. (2) Guava: the student lifts a cup to reveal the hidden cube, sets the cup aside and picks up the cube. (3) Guava: notices a knocked-off bowl and restores it. (4) SimplerEnv Google robot (zero-shot): done is rejected at 11.9 cm, below the student’s own 12 cm check; it lifts to 15 cm and finishes. (5) EB-Manipulation (zero-shot): “rolling pin” is not found, so the student describes it by shape (“black cylinder”).
Figure 5: Choosing the right object, the untrained baseline, and a failure. (1) LIBERO-PRO, language perturbation: of two identical black bowls, the student picks the one on the cookie box after inspecting close-ups. (2) The untrained Qwen3.5-4B in H∗ on the episode of Fig. 3 , row 5, repeats grasp(eggplant) 22 times ( NOT_SEEN ) until the budget runs out; the student re-describes the object and succeeds in 7 calls. (3) SimplerEnv Google robot (zero-shot), same episode and harness: distilled vs. untrained 4B. (4) A student failure on RoboTwin: the shoe ends OFF_TARGET by 2.3 cm, the student’s own checks pass and it declares success, but the simulator judges a failure.
Harness
H1
H2
H7
H12
H15
H16
H∗
Success
33.4
33.6
49.4
50.8
54.6
60.4
63.6
Appendix
Table 8: Accepted harness snapshots. Teacher success (%) on the 500 RoboTwin-50 evaluation episodes (fixed denominator) for accepted harness snapshots; Hk is the k -th accepted snapshot and H∗ the frozen final version.
Language model agents are deployed with a harness, the software around the model that manages its context, tools, and feedback. When such an agent is distilled into a smaller one, the harness stays in place, so the student mainly needs the teacher-specific abilities that the harness cannot provide, such as acting correctly on harness information. Standard distillation, however, imitates the teacher's full outputs and treats the harness as part of the input. We propose Harness-Aware Distillation (HAD), which focuses distillation on what the teacher adds beyond the harness. HAD complements on-policy distillation with two components: an action preference that contrasts the same teacher's actions with and without the harness information, scored after the student's own reasoning, and a validity check that drops preference pairs whose preferred action contradicts the harness records. We show that the contrast gives the student information that imitating the teacher alone cannot provide, and HAD needs no task rewards, success labels, or future information. Across multiple long-horizon agent benchmarks and models, HAD outperforms on-policy distillation baselines with the same fixed harness. Our analysis shows that HAD enters fewer unproductive loops and recovers from errors more often than the baselines, and suggests that it adaptively keeps learnable feedback in its weights while reading state information from the harness.
A central goal in robotics is to enable manipulation across changing tasks and environments. Vision-language-action (VLA) models provide broad manipulation capabilities but can struggle when execution requires diagnosing failures and adapting behavior. Strong agents can discover effective interventions through interaction with these policies. We propose Recursive Harness Distillation to accumulate this experience as reusable guidance across agents. A strong agent distills its experience into a playbook for a light agent, then recursively refines the playbook using the light agent's execution feedback. The resulting playbook enables agents to reuse accumulated intervention knowledge in new task instances without updating model parameters. In real-world manipulation, the harness improves success from 37.3% to 64.0%. On SimplerEnv Bridge, the light agent with the playbook achieves 66.7% success, compared with 41.7% for the GR00T-only baseline, and outperforms the strong agent without a playbook. The same playbook also benefits the strong agent, which reaches 79.2% success. These results demonstrate the feasibility of harness distillation for robotics: intervention experience can be accumulated, refined through execution, and reused across agents to improve manipulation.
Seungyeon Kim, Junhoo Lee, Minkyu Kim +2
Seoul National University · Korea Advanced Institute of Science and Technology
An LLM agent is defined by two things: the weights inside its model and the harness of components assembled around it. Harnesses are still handcrafted, and HarnessX, which evolves them automatically, starts each benchmark from a handcrafted harness, reports gains on the tasks it evolved on, and budgets 100 to 175 million meta-agent tokens per benchmark. We propose LiteEvo, a lightweight harness-evolution algorithm whose tool-free meta-agents mine agent trajectories for reusable components, curate them into a versioned library, and compose each round's harness from it, starting every benchmark from the same neutral harness and never naming the benchmark. Evolving on the graded tasks of five agentic benchmarks with a frozen Qwen3.5-9B, LiteEvo lifts pass@2 by 10.5 to 67.7pp and reaches comparable or higher pass@2 than a reproduction of HarnessX (71.0 against 67.3 on average) at 13.0 lower mean API cost. Harnesses evolved on train tasks keep their gains on unseen test tasks of four benchmarks, and LiteEvo also lifts Claude Code with Sonnet 4.6 by 1.2 to 71.4pp.