Source-linked AI summary
RSI-Anything: Unlocking Recursive Self-Improvement with World Models
Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao, Wanli Ouyang, Pheng Ann Heng, Zhenfei Yin, Yingcheng Wu, Ling Yang
TL;DR
Recursive self-improvement must turn interaction into persistent progress despite costly external feedback and prediction reliability challenges as policies evolve. RSI-Anything uses predictive world models to guide decisions and learning, with feedback-driven updates and retention checks. Across computational, planning, and scientific settings, the evaluations report gains from predictive selection, adaptation, and retained updates.
Problem
Recursive self-improvement must carry interaction experience into later rounds, but external feedback is costly and policy changes can undermine world-model prediction reliability.
Source: §1 · p. 2Method
RSI-Anything uses world-model predictions to guide actions, evaluations, and learning-data selection, while feedback drives policy or workflow updates and conditional model adaptation.
Source: §1 · p. 3Results
Across its evaluated settings, RSI-Anything reports successive gains from predictive selection and model adaptation in code repair, improved laboratory planning across retained policy versions, and higher GUI task completion.
Source: §1 · p. 3, §5 · p. 19Takeaways & Limitations
The results connect prediction-guided decisions with feedback-driven learning and retained system changes across the evaluated settings.
Source: §5 · pp. 19–20Joint multiround evaluation is needed to establish sustained mutual improvement of policy and world model within a single task.
Source: §5 · p. 20
Abstract
from arXiv · showhide
Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection. Observed feedback drives policy or workflow improvement and conditional world-model adaptation, while component and compatibility checks govern the retention of updated policy-model pairs. Across diverse computational and scientific domains, RSI-Anything demonstrates substantial gains. In code repair, with the coder fixed, success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation, at comparable test usage. Graphical user interface (GUI) evaluations showed higher task completion, while average actions decrease by 39.4%. Across five successive policy versions, laboratory-planning scores more than double from 21.80% to 46.01%. In reaction optimization, RSI-Anything achieves 96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget. An end-to-end immune-science application further connects workflow evolution and predictive modeling with therapeutic hypothesis generation. These results demonstrate how predictive world models and retained system updates enable recursive improvement across diverse environments.
1. Introduction
The introduction frames recursive self-improvement as carrying interaction experience into persistent improvements across rounds, despite costly feedback and prediction reliability challenges as policies evolve. RSI-Anything combines predictive guidance with feedback-driven updates and retention checks, and is evaluated across computational, planning, and scientific settings.
1. Introduction
- Reliable recursive improvement must address costly external feedback and the coupled effects of policy and world-model updates on prediction reliability and later performance.Policy changes alter candidate actions and experience distributions, while world-model changes alter candidate rankings and learning signals.Source: §1 · p. 2
- RSI-Anything is a domain-general framework that uses predictive world models to guide action selection, external evaluations, and learning-data selection or construction.Feedback drives policy learning or workflow revision and triggers world-model adaptation when reliability checks warrant it.Source: §1 · p. 3
- Evaluations span code repair, GUI interaction, physical-domain planning, experimental science, and an immune-science application involving workflow evolution and therapeutic hypothesis generation.The evaluations examine predictive candidate selection, retained policy or model updates, and performance under limited external feedback.Source: §1 · p. 3
- Component checks assess individual updates, while compatibility checks select which policy–model pair is retained to shape subsequent candidates and learning experience.The components may evolve asynchronously, and accepted revisions feed into later rounds.Source: §1 · p. 3
2. RSI-Anything Framework
RSI-Anything uses world-model predictions to guide actions and feedback collection, then uses experience to propose and retain system updates that shape later interactions. Its framework combines prediction, feedback-driven improvement, and checks on individual components and their assembled policy–model pair.
2. RSI-Anything Framework
- The retained policy–model pair shapes subsequent candidates and learning experience, connecting prediction-guided selection, feedback-driven updates, and system retention into a recursive loop.The framework can instantiate different models, action encodings, outcome readouts, and pathways across domains.Source: §2 · pp. 3, 5, Figure 2 · p. 4
2.1. Problem formulation
- Within a run, task objectives, interfaces, scoring criteria, constraints, rules, and total budgets remain fixed, while task states, feedback, candidate pools, and accepted versions can evolve.An update cycle may aggregate multiple actions, trajectories, or tasks, and counts cycles rather than successful revisions.Source: §2.1 · p. 5
2.2. Predictive foundation
- JEPA-Anything’s orthogonal predictive factorization organizes latent targets into complementary factors, predicts them from context, and synthesizes a complete latent state.A readout converts predicted latent targets into outcome predictions; one-step selection needs only an outcome readout, while multistep planning requires a state-propagation adapter, and rolled-out states remain unverified until checked externally.Source: §2.2 · pp. 6–7
2.3. Prediction-guided selection
- Candidate selection predicts outcomes or bounded rollouts, excludes invalid candidates, and chooses feasible candidates for feedback under the remaining budget.When enabled, selection may penalize uncertainty or prioritize uncertain candidates; imagined training data is governed by a separate admission filter and evidence criteria.Source: §2.3 · pp. 7–8
2.4. Feedback-driven improvement
- External feedback supports separate policy-improvement and predictor-adaptation data, with policy learning or workflow revision and world-model adaptation conditional on reliability, evidence, and budget.Reliability assessment uses predictions attributable to the current model version; its conclusions are limited to the observed and covered candidate range, and insufficient coverage does not activate adaptation.Source: §2.4 · pp. 8–9
2.5. Recursive retention
- Retention checks evaluate proposed components and their assembled pair, with pair acceptance testing system decision behavior and predictor suitability on candidates proposed by the selected policy.Failed component checks retain the current component, while a failed pair check retains the current pair; accepted revisions shape subsequent candidates, rankings, and learning signals.Source: §2.5 · pp. 9–11
3. Results
The results examine world-model-guided selection and adaptation across code repair, GUI interaction, laboratory planning, reaction optimization, and an immune-science workflow.
3.1. World-model-guided selection and adaptation improve code repair
- 31.2% hidden-test success with world-model adaptation followed 29.2% with fixed-world-model selection and 26.3% with a static verifier, using the same coder.Online test-suite usage was similar across the static-verifier and fixed-world-model configurations, at 3.72 and 3.71 calls per task, respectively.Source: §3.1 · p. 11, Figure 3 · p. 12
3.2. RSI-Anything improves task completion in GUI environments
- 99% final click-checkboxes success accompanied a 39.4% reduction in mean actions per episode from baseline; OSWorld success gains were 18.4% relative to the initial agent.The click-checkboxes results also report repeated clicks falling from 237 to 0 and incorrect-submission episodes from 9 to 1.Source: §3.2 · p. 13, Figure 4 · p. 14
3.3. Successive policy updates improve laboratory planning
- The LabHorizon Final Score rose across five retained policy versions from 21.80% to 46.01%, a 24.21-percentage-point increase from the initial version.The final 4B version scored 46.01% on Level 2 versus 45.32% for the Actor–Simulator–Selector reference.Source: §3.3 · p. 14, Figure 5 · p. 15
3.4. Prediction- and feedback-guided search improves reaction yields at matched experiment counts
- After 40 experiments, mean best observed yield was 96.15% for RSI-Anything versus 82.16% for Bayesian optimization in the flow LC–MS screen.In the separate 20-experiment GC screen, best observed yield was 93.0% versus 87.5%, respectively.Source: §3.4 · p. 16, Figure 6 · p. 17
3.5. An end-to-end application of RSI-Anything in immune science
- The immune-science application combined a retained model-building workflow and an action-conditioned immune world model with evidence-guided analysis, culminating in a prospective SIRP𝛼-inhibition and IL-36𝛾-axis hypothesis [13].The analysis used model forecasts and experimental evidence to rank interventions and revise hypotheses.Source: §3.5 · p. 16, Figure 7 · p. 18
4. Related Work
Related work spans recursive agent improvement, feedback-based policy learning, predictive world models, and scientific or laboratory agents. RSI-Anything connects these lines through predictive action selection and retention of compatible policy–model pairs.
4.1. Recursive self-improvement
- RSI-Anything extends recursive-improvement approaches that update agent code, policies, memories, or training environments by explicitly incorporating predictive world models and retaining compatible policy–model pairs.Prior examples include CURE [14], Voyager [32], and RLAnything [21].Source: §4.1 · pp. 18–19
4.2. Verification and policy self-improvement
- Feedback-based methods such as CURE [14] and RLAnything [21] motivate policy improvement, while RSI-Anything makes predictor reliability and policy–model compatibility explicit in update acceptance.Changing the policy also changes the candidate distribution on which its predictor must be evaluated.Source: §4.2 · p. 19
4.3. Predictive world models
- Predictive environments support control and planning [6] [8], while RSI-Anything organizes predictions and observed outcomes to guide action selection and acceptance of later system versions [9].Source: §4.3 · p. 19
4.4. Scientific and laboratory agents
- Scientific-agent work supplies domain capabilities including LabHorizon, reaction prediction, and cellular-response forecasting; RSI-Anything connects them through a common transition-and-feedback contract.Source: §4.4 · p. 19
5. Discussion
Results support prediction-guided feedback selection and predictor improvement as complementary roles in recursive improvement. The discussion links prediction–observation pairs to retained system changes while noting boundaries on the evidence.
5. Discussion
- Across code repair, GUI tasks, and laboratory planning, results show successive code gains, higher GUI completion, and improvement across five policy versions.The discussion also connects measured-feedback search and the immune-science workflow to learning and retained system changes.Source: §5 · pp. 19–20
- Prediction–observation pairs connect current action selection with assessment of system changes for subsequent decisions.Policies and world models may change on different schedules, with reliability checks guiding model adaptation and component and compatibility checks governing retention.Source: §5 · p. 20
- Domain generality means reusing the improvement mechanism across domain-specific models, policies, and evaluations; sustained mutual improvement requires same-task multiround evaluation.The laboratory results assess protocol-aligned planning, chemistry results cover two Suzuki–Miyaura screens at reported budgets, and the immune-science therapeutic hypothesis awaits experimental testing.Source: §5 · p. 20
6. Experimental Methods
The experimental methods instantiate a shared framework across distinct domain settings, models, prediction targets, and feedback pathways. They evaluate code repair, GUI interaction, laboratory planning, reaction optimization, and an immune-science application, with statistical reporting tailored to each setting.
6. Experimental Methods
- Domain settings bind different states, actions, prediction targets, and feedback sources to the shared framework; Table 4 summarizes the domain-specific models and update pathways.Experiments evaluate predictive selection, predictor adaptation, successive policy versions, GUI policy and world-model updates, feedback-guided chemistry search, and immune-science analysis with a fixed predictor.Source: §6 · p. 20, Table 4 · p. 21
6.1. Domain-specific evaluation settings
- Code repair compares predictive selection and world-model adaptation with random selection and a static verifier, using a fixed coder on CodeContests and additional APPS and LiveCodeBench evaluations.CodeContests uses 200 tasks, three seeds, and at most three feedback rounds; APPS and LiveCodeBench use shared initial programs, four candidates per round, and at most three repair rounds, with full-test success as the endpoint.Source: §6.1 · pp. 20–22
- GUI evaluations use MiniWoB++ click-checkboxes and OSWorld, with interaction outcomes updating both the action policy and world model.The checkbox evaluation tracks success, actions, repeated clicks, and incorrect submissions; OSWorld compares the initial agent, RLAnything, and RSI-Anything.Source: §6.1 · p. 22
- LabHorizon F200 assesses next-action accuracy and protocol-aligned action sequences across five consecutive 4B policy versions, alongside an Actor–Simulator–Selector reference.Level-2 Final Score combines action-sequence similarity and parameter accuracy.Source: §6.1 · p. 22
- Reaction optimization uses a ReactionT5–JEPA-Anything adapter to predict yield distributions and guide condition choices in two screens, with Bayesian optimization as baseline.The screens use different catalyst, base, and solvent settings and report best observed or mean best observed yield at matched experiment counts.Source: §6.1 · p. 22
- The immune-science application evolves scientific-agent workflows and constructs an action-conditioned immune model for forecasting, cross-scale prediction, and action composition.Downstream intervention analysis freezes the resulting model parameters while the analysis state, candidate interventions, and hypothesis ranking evolve through inference and self-correction.Source: §6.1 · pp. 22–23
6.2. Statistical reporting
- Statistical reporting uses task-level clustered inference for 600 CodeContests evaluations, three-seed means for APPS and LiveCodeBench, 200-task LabHorizon means, and running-best chemistry yields at matched experiment counts.The reported 24.21- and 0.69-percentage-point gains are differences between benchmark means.Source: §6.2 · p. 23