Skip to main content
Search

# Recursive Self-Improvement

27 related articles

Subscribe to updates
  1. arXiv

    Reglos splits personal-agent memory into two loops: memory grows with each chat while a coding agent rewrites the harness

    Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.

  2. arXiv

    Language agents get a second trainable policy: JoMoHa learns the harness too, averaging 70.3% success across three environments

    A language agent's performance depends jointly on its model weights and on the harness, the runtime program that organizes how those weights are used, so optimizing either alone is constrained by the other. JoMoHa trains a harness-generating Engineer and a task-executing Solver in one reinforcement-learning process, using the same task outcome to update both policies. It reaches 97.0% success on ALFWorld, 45.8% on WebShop and 68.0% on DBBench, an equally weighted average of 70.3%, above the evaluated weight-only, harness-only and alternating-training recipes.

  3. arXiv

    Agents no longer accept self-edits on a higher score alone: TRUST cuts accepted bad updates from 45.59% to 2.78%

    Agents can rewrite their own prompts, memory, tools and workflows without retraining the underlying LLM, but the common greedy gate accepts any update whose validation score rises, so sampling noise can pass as real progress. TRUST recasts each round's acceptance as an online multiple hypothesis test, uses an LLM over historical context to set how much evidence a candidate needs, and corrects for reused validation samples. Across six settings on HotpotQA, DROP and APPS with DeepSeek-V4-Flash and Qwen-3.8-Flash, the mean share of accepted inferior updates falls from the greedy gate's 45.59% to 2.78%, and mean maximum drawdown falls from 4.963 to 0.047 percentage points.

  4. arXiv

    Agents no longer learn from a scalar reward alone: ACE-RL trains the reflections and notes they hand forward, topping held-out reward on four benchmarks

    Agent reinforcement learning usually compresses a trial into one scalar reward, discarding the error logs and tool outputs the environment revealed; ACE-RL also trains the reflections and task notes the model writes, scoring each candidate note by a fresh attempt that executes it. At equal training compute on Mastermind, MBPP, AppWorld and MiniSWEBench, ACE-RL's held-out reward is higher than actor-only GRPO, ERL, and delayed-joint PPO trained end to end through the harness with a learned critic.

  5. arXiv

    Agents must now identify the PDE themselves: RSI-PDEbench tests them on 1,001 cases, lifting mean pass rates by up to 181.9% across four models

    Recovering a partial differential equation from sampled solution data demands expert work in equation discovery, coefficient estimation and numerical solver configuration. RSI-PDEbench turns that workflow into a benchmark of 1,001 cases over 42 equation templates, split into noise-free, 50% Gaussian-noise and unseen-template-with-shifted-coefficient settings; the accompanying RSI-PDE agent shortlists candidate equations by residual, refines coefficients and solvers with relative feedback, and keeps improvements behind a validation gate. Across four frontier models its mean pass rate rises 106.5–181.9% over single-run baselines and 3.0–19.6% over native agent harnesses.

  6. arXiv

    Clinical speech sharing drops one-size-fits-all anonymisation: CASA tunes per task and clears privacy and diagnosis on all 14 tasks

    Clinical speech carries both speaker-identifying voiceprints and sensitive spoken content, while different clinical tasks rely on different speech cues, so a fixed anonymisation pipeline struggles to keep both privacy and diagnostic value. CASA recasts anonymisation as an optimisation driven by a task description, repeatedly evaluating and adjusting strategies across voiceprint transformation and sensitive-content masking, and storing cross-task experience in memory. Across 14 clinical tasks in five languages, CASA met speaker-anonymisation and downstream classification requirements on all of them, versus 8 for the strongest fixed method and 4 for SecureSpeech.

  7. arXiv

    Remote sensing models stop being hand-built: given a harness and a sealed scorer, an agent beats published results on four benchmarks

    Remote sensing models have long been built by hand, with experts designing a training recipe for each new task and tuning it against a benchmark. RSI4EO casts a task as two parts: a harness that fixes the data, pretrained weights and compute an agent may use, and a sealed verifier that returns only a score. A general-purpose coding agent then writes the training code, trains models and improves them recursively. On four benchmarks with published results, spanning optical, SAR and multispectral time-series imagery, the agent-built models surpass the state of the art, and with about 1% of the training labels they keep most of that accuracy.

  8. arXiv

    Recursive self-improvement reaches frontier pre-training: 801-30B-A3B trains from scratch in three months and ranks second among base models

    Recursive self-improvement had only been closed on post-training, where each trial costs GPU-hours and returns a verifiable score; a frontier pre-training run costs hundreds of thousands of GPU-hours, cannot be repeated, and has no single-score objective. A newly formed team at Shanghai Innovation Institute and Xinxu AI split pre-training into smaller closable loops, using agents for debugging, tiny-SFT and related measures to estimate post-training headroom, and quantified workflows for data tasks, producing the 30B-A3B base model 801 in three months. On 13 base benchmarks it averages 73.13, behind only the larger Qwen3.5-35B-A3B at 76.67.

  9. arXiv

    Recursa lets a foundation model system rewrite itself: all-pass on held-out legal tasks rises from 30.0% to 55.0%

    Recursa splits adaptation to a new domain into two steps: it bootstraps a navigable knowledge repository from domain materials, then has the same model act as both solver and learner, with parallel child learners diagnosing, editing and validating revisions to instructions, skills and memory. On Harvey LAB's Firm Knowledge legal benchmark, all-pass rate on held-out test tasks rises from 30.0% to 55.0% with DeepSeek-V4.1-Flash as solver and from 35.0% to 46.0% with Qwen3.8-Flash, with no stronger model or human expert writing the revisions.

  10. arXiv

    RSI-Env lets an agent rewrite both model and harness, then hands the next round to the accepted version, with gains even without benchmark feedback

    RSI-Env places model training and harness modification in one persistent sandbox, with an independent Host Controller handling services, validity checks, versioning and capability evaluation, and in Self mode each accepted model–harness version takes over the next improvement round. Starting from Qwen3.8-27B with a minimal harness, Open-General Self, which receives no benchmark feedback, raises GPQA Diamond accuracy from 65.2% to 73.2% and AIME 2026 from 53.3% to 56.7%; Knowledge Self lifts SimpleQA from 82.0% to 85.0% and then 88.0%, while the same runs fall back on other benchmarks.