Pranay Gupta — direct weight manipulation of π0.5 beats fine-tuning on cloth folding (50% → 68% FT → 75% weight edit)
Pranay Gupta (Argon Robotics, previously CoRL’26) runs the first filed robotics transfer of the Weight-Space Programming paradigm (Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights) onto a VLA: cloth-folding on π0.5 evaluated on RoboDojo, comparing base checkpoint vs fine-tuning on GPT-6 Astra corrections vs direct weight manipulation. The reported ordering is Base 50% → FT on Astra corrections 68% → direct weight manipulation 75%. Full writeup promised; tweet is the only public artifact at filing time. If the result holds, this is the first datapoint that the no-gradient weight-editing recipe Takehi et al. demonstrated on Qwen3.5-0.8B transfers to a flow-matching action-expert VLA on a manipulation task.
Key claims
Section titled “Key claims”- Base π0.5 scores 50% on cloth-folding evaluated on RoboDojo [tweet body].
- Fine-tuning π0.5 on “Astra corrections” scores 68% on the same task and benchmark [tweet body].
- Direct weight manipulation of π0.5 (no fine-tuning) scores 75% on the same task and benchmark [tweet body].
- Author pitches this as a direct robotics analog of Takehi et al.’s Weight-Space Programming result on Qwen3.5-0.8B (Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights) — the quoted tweet in the thread [tweet body, quoted Rikiya Takehi tweet].
Method
Section titled “Method”Tweet-only disclosure; the method behind the “direct weight manipulation” column is not specified. The quoted-tweet lineage points at Takehi et al.’s recipe (ask a frontier LLM with Python sandbox access to read weights, propose and evaluate edits over a budget of forward passes, keep only edits that improve a measured outcome, no gradients) applied to π0.5’s weights with a cloth-folding task-success metric. What “Astra corrections” means for the fine-tuning arm is also undefined — plausibly either corrective trajectories collected under an Astra-driven harness or Astra-authored DAgger-style corrections, consistent with Argon’s “Chief Robot Officer” pitch.
Sample size, number of rollouts per cell, which π0.5 checkpoint, whether the eval is RoboDojo-Sim or a specific task subset, and the forward-pass budget for the weight-edit arm are all unreported. Flagged for /bud refresh once the full breakdown is published.
Results
Section titled “Results”Single row of three numbers:
| Treatment | Success |
|---|---|
| Base π0.5 | 50% |
| Fine-tuned on Astra corrections | 68% |
| Direct weight manipulation | 75% |
No error bars, no rollout counts, no per-subtask split. The 7 pp gap between fine-tuning and weight manipulation is the headline; the 18 pp gap over base at zero gradient cost is the interesting one if it holds.
Why it’s interesting
Section titled “Why it’s interesting”Direct extension of Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights from a small LLM onto a production flow-matching VLA on a contact-rich manipulation task — the first filed datapoint for whether the “LLM edits weights without gradients” recipe transfers across the model-class gap (LLM → VLA) and the benchmark gap (MATH-500 / GSM8K → cloth folding on RoboDojo). Takehi et al.’s own null result on Gemma 3 270M (editing removes obstacles that hide an ability but cannot add ability the model lacks) makes this result surprising: cloth folding on π0.5 is not obviously a “remove the self-locking reasoning-loop pathology” situation of the kind that let Astra lift Qwen3.5-0.8B from 7.9% to 48.1%. If the 75% figure replicates, it argues either (a) π0.5’s base policy also has an obstruction-removable pathology on cloth folding, or (b) the ability-vs-obstruction dichotomy from the Qwen result doesn’t carry over to VLAs cleanly.
Direct counter-recipe to the fine-tuning entries on the VLA Models recipe-lever board: Dream Machines’ π0.5 fine-tuning map on bimanual pick-and-insert (Empirical results fine-tuning π0.5 on a real manufacturing task) shows LoRA at rank 64/256 underperforms full fine-tuning by 23–43 pp with the author’s hypothesis “new embodiment is not a low-rank edit.” Gupta’s result cuts the other way: a (presumably) low-fraction gradient-free edit beats a full FT baseline. The reconciliation question — is cloth-folding specifically obstruction-removable where embodiment transfer is not? — is the useful one to put in the next version of the full writeup.
Also worth flagging the direct connection to Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates‘s finding that full-weight VLA fine-tuning in bfloat16 without an fp32 master copy silently erases most updates. If Gupta’s “fine-tuned on Astra corrections” baseline is running under this pathology, the 68% number may be a low estimate rather than a fair comparison for the fine-tuning arm.
See also
Section titled “See also”- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights — the originating paradigm; Gupta’s tweet is a quote-retweet of the Takehi thread
- VLA Models — concept home; adds a new row to the recipe-lever board (gradient-free LLM-authored weight edit on top of a frozen base VLA)
- Empirical results fine-tuning π0.5 on a real manufacturing task — sibling empirical π0.5 fine-tuning map that reports LoRA underperforming full-FT; direct counter-pattern on which edit shape works for VLAs
- Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates — bfloat16 full-weight VLA fine-tuning silently erases most updates — a candidate reason the fine-tuning baseline may be weaker than it looks
- AI-for-AI Research — the natural outer-loop concept (LLM-as-optimizer at the weight granularity, applied here to robotics)
- Parameter-Efficient Finetuning — if the weight-edit arm modifies a small fraction of π0.5, it lands as a gradient-free competitor to LoRA on the same task
- Mechanistic Interpretability — Takehi et al.’s diagnosis → intervention loop in a VLA setting