Alphabell.
Paper · Tools and harness

Recursive Self-Improvement on a Budget

Recursive Self-Improvement on a Budget: Scoring Code Edits Before They RunA coding agent uses a linear readout to score proposed edits to its own scaffold before running them, improving its solve rate on held-out instances.

Recursive Self-Improvement on a Budget
Figure 1 of the paper, arXiv. Source
The loop

A coding agent proposes and scores edits to its own scaffold, which improves its solve rate. This higher solve rate then guides the selection of the next scaffold edit.

The loop
Recursive Self-Improvement on a Budget
Paper · arXiv
  1. Agent proposes scaffold edits
  2. Agent scores edits before running
  3. Agent selects highest scoring edit
  4. Agent uses improved scaffold
  1. Agent proposes scaffold edits
  2. Agent scores edits before running
  3. Agent selects highest scoring edit
  4. Agent uses improved scaffold
↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

Scoring proposed edits before executing them drastically reduces the evaluation budget required for open-ended self-modification.

Evidence

Improves solve rate on held-out HumanEval+ and Polyglot instances compared to the root scaffold.

arxivscaffold-evolutionbudget-efficiency

Related loops

More tools and harness →
ScholarEvolve
PaperarXiv
ScholarEvolve
  1. Research agent writes harness mutation blueprints
  2. Coding agent implements harness module changes
  3. Winning harness becomes the new champion
  4. New champion harness runs further generations
  5. repeat
Tools and harnessYang et al. (UC Santa Barbara, Microsoft) · 2026
Training the Whole Agent, Not Just the Model
PaperarXiv
Training the Whole Agent, Not Just the Model
  1. Teacher models evolve tools and skills
  2. Policy trains with adapted harness
  3. Policy generates new trajectories
  4. Trajectories guide further harness adaptation
  5. repeat
Tools and harnessLingxiao Tang et al. · 2026
Self-Evolving Coding Agents via Harness-Policy Optimization
PaperarXiv
Self-Evolving Coding Agents via Harness-Policy Optimization
  1. Policy analyzes its own failures
  2. Proposes new harness candidates
  3. Policy trains via RL on new harness
  4. Successor policy tackles harder tasks
  5. repeat
Tools and harnessZhepei Wei et al. · 2026