WeightLoop: Weight-Space Programming with Experimental Feedback

Mengran Li, Bo Li, Weijing Zhao, Chengyang Zhang, Boyu Zhang, Jinfeng Xu, Zelin Zang, Victor C. M. Leung, Stan Z. LiarXiv 2026 · arXiv

TL;DR

WeightLoop is a novel framework that uses Large Language Models (LLMs) to write specialized "editor" programs capable of surgically updating a deployed model's weights based on new experimental feedback. By turning weight-space programming into a continuous loop, it enables persistent model improvement without the need for expensive gradient retraining or repeated LLM calls.

What Problem Does This Paper Address?

In scientific discovery, we often use predictive models to rank which experiments to run next. However, once an experiment is completed, we face a bottleneck: how do we incorporate that fresh, high-quality data back into the model?

Traditional approaches like full retraining are computationally expensive and often require more data than a single batch of experiments provides. Fine-tuning can be brittle and prone to catastrophic forgetting. The authors identify a gap in "experimental decision-making" where the predictor remains static while the data evolves. They propose a way to make experimental evidence a direct, automated driver of model change, ensuring that subsequent decisions immediately reflect what was just learned.

Methods

The core of the paper is WeightLoop, a framework inspired by Recursive Self-Improvement (RSI) and Weight-Space Programming (WSP). The workflow consists of two distinct phases:

  1. One-time Program Construction: An LLM (specifically GPT-6 Astra) inspects the task objective and the model's implementation. It then writes a "bounded, reusable parameter-editing program"—the editor. This editor defines which parameters are mutable, how to calculate candidate updates, and how to validate them against new data.
  2. Repeated Execution: Once the editor is written, the LLM is no longer needed. As new experimental measurements arrive, the editor identifies discrepancies between predictions and outcomes. It then constructs and validates candidate weight edits (e.g., attenuating specific connections in a neural network or adjusting leaf values in a decision tree) and writes the selected changes directly back to the model checkpoint.
Figure 1: WeightLoop separates one-time program construction from repeated execution with experimental feedback. The LLM builds a reusable editor; returned measurements validate parameter changes, followed by model save, reload, and the next decision. Weight values and ranking gains illustrate a HepG2 example.

Crucially, this process happens via numerical execution without backpropagation or gradients, preserving the original inference interface of the deployed model.

Datasets

The authors evaluated WeightLoop on several real-world biological discovery tasks using public functional screens:

  • DepMap (Cancer Dependency Map): Used for initial editor design and testing single-round ranking (HepG2 and Jurkat cell lines).
  • Cellular Physiology Screens: HL60 proliferation and migration assays.
  • Autophagic Flux: Measuring LC3 and NBR1 protein levels.
  • LDL Uptake: Measuring reduced LDL uptake in HuH7 cells.

These datasets provide measurable, concrete feedback (e.g., "how much did this gene perturbation affect cell growth?") that the editor uses to refine the model.

Key Insights and Ideas

  • Surgical Precision: One of the most striking findings is that significant performance gains can be achieved by editing a tiny fraction of the model. In one example, the editor changed only 16 weights out of over 3.6 million (0.00044%) to improve discovery utility.
  • Fixed Editor, Evolving Predictor: By freezing the editor code, the system avoids the instability of "models editing themselves" indefinitely. The LLM provides the logic for improvement, while the experimental data provides the signal.
  • Interface Agnostic: The paper demonstrates that this "editor" concept works across different model architectures, including Neural Networks (MLPs), Ridge Regression, and XGBoost.
  • Persistence: Unlike simple "in-context learning" where an LLM might just look at a few examples to make a better guess, WeightLoop creates a persistent update to the model's weights that survives checkpoint reloads.

Key Results

WeightLoop consistently outperformed both the original un-edited models and "random edit" controls across various settings:

  • Single-Round Ranking: In seven out of eight endpoint-budget settings, targeted edits improved the mean ranking utility (how well the model ranks the best candidates).
  • Sequential Discovery: In tasks like autophagic flux (LC3/NBR1) and LDL uptake (HuH7), the "Edited" policy consistently found more high-value genes over multiple rounds of acquisition compared to the original model.
Measured feedback changes discovery and prediction. (a–c) Complete acquisition curves show the "Edited" (blue) line consistently outperforming the "Original" (dashed grey) and "Permuted" (dotted brown) controls.
  • Quantitative Gains: For sequential discovery in the LC3 task, the edited model showed a discovery utility increase ($\Delta D_{100}$) of +0.0334 over the original. In the HepG2 growth task, the "Bottom-Up" utility at budget 20 (BU@20) rose from 0.6046 to 0.6213.

Limitations and Future Work

  • Bounded Edits: The current implementation focuses on "readout" layers or leaf values. While effective, it doesn't yet explore editing deep internal representations which might be necessary for more complex task shifts.
  • LLM Dependency for Setup: The quality of the editor depends entirely on the LLM's ability to reason about the provided code and task. If the model architecture is highly unconventional, the LLM might struggle to write a valid editor.
  • Future Directions: The authors suggest that future work could involve more complex validation rules or allowing the editor to suggest new features to be learned, rather than just reweighting existing ones. There is also potential to apply this to non-biological domains where experimental feedback is expensive but valuable, such as materials science or hardware optimization.

Summary updated

Abstract

Weight-space programming (WSP) shows that language models can reason about and rewrite neural network weights; recursive self-improvement (RSI) motivates feedback-driven improvement. Models guiding experiments receive concrete feedback: completed assays reveal prediction errors and inform what to test next. We introduce WeightLoop, a framework that turns experimental feedback into persistent updates to deployed predictors. GPT-6 Astra inspects a task specification and the model implementation, then writes a bounded, reusable parameter-editing program. As new measurements arrive, the program identifies prediction discrepancies, validates candidate edits using observed outcomes, and writes the selected changes back to the original model. The editor remains fixed while the predictor evolves, enabling repeated updates without further LLM calls or gradient retraining while preserving the original inference interface. In offline replay, targeted edits improve mean ranking utility in seven of eight endpoint-budget settings (all eight versus random edits) and mean final discovery across three sequential tasks. WeightLoop connects WSP with an RSI-inspired improvement loop, making experimental evidence a direct driver of model change and subsequent decisions.
BibTeX
@misc{li2026weightloop,
  author = {Mengran Li and Bo Li and Weijing Zhao and Chengyang Zhang and Boyu Zhang and Jinfeng Xu and Zelin Zang and Victor C. M. Leung and Stan Z. Li},
  title = {{WeightLoop: Weight-Space Programming with Experimental Feedback}},
  year = {2026},
  eprint = {2611.07319},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.07319},
}
Figure