Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.07495
Learning Multi-Agent Orchestration:Higher-Order Interactions in Policy Improvement
Learning Multi-Agent Orchestration introduces Interaction-Aware Policy Optimization (IAPO), an approach for training an LLM orchestrator whose decisions influence one another. Instead of crediting each worker call independently, IAPO decomposes policy improvement into individual-decision and pairwise interaction terms, fits bounded candidate updates, and validates them on fresh trajectories before deployment. The paper evaluates long-input versions of GSM8K, MATH, DROP, MBPP, and HumanEval with Qwen2.5-7B-Instruct workers. IAPO reaches or exceeds the strongest baseline at matched token budgets while using substantially fewer tokens per episode.
How it works
The environment is an episodic MDP with tasks q drawn from a distribution, a set of workers, a finite output space, and at most K calls. At an orchestrator step, the action is either stop or a pair (j,E), where j identifies the worker and E is a subset of earlier outputs made visible to that worker. The worker then samples an output from its fixed policy, the output is added to the history, and control returns to the orchestrator. A call to worker j receives reward −lambda c-j, while stopping receives the task reward R(q,h).
What the paper evaluates
The empirical evaluation has two parts. First, controlled tasks with enumerated trajectories test whether interaction-aware scoring recovers gains missed by first-order scoring and avoids harmful updates. The tasks are Complementarity, Redundancy, and Evidence threshold. Second, the paper evaluates LLM orchestration on constructed long-input versions of GSM8K, MATH, DROP, MBPP, and HumanEval. Each target problem is placed among 30 to 90 units from other problems; the resulting documents contain 81, 91, 31, 61, and 61 units respectively. The test splits contain 790, 220, 800, 333, and 131 problems. These are explicitly constructions rather than the standard benchmark distributions.
Limit to keep in view
The benchmark inputs are constructed long-input variants rather than the original GSM8K, MATH, DROP, MBPP, and HumanEval distributions, so the absolute accuracies should not be compared directly with standard benchmark reports.
arXiv:2610.39714
MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery
MolRSI is a molecular-discovery agent that improves how it builds computational workflows by retaining experience across tasks. It converts literature and code into validated executable skills, uses experience-guided Monte Carlo tree search to assemble and refine workflows, and stores workflow performance, modification outcomes, and execution repairs in persistent memory. Across antibiotic activity prediction and peptide–MHC binder design, cross-task experience improved workflow selection and produced stronger prospective results: 17 of 107 screened compounds met the MSSA hit criterion, compared with 21 of 241 for a prior expert workflow, while pMHC campaigns produced eight experimentally validated, preferential binders.
How it works
MolRSI uses an Explore Agent to retrieve relevant literature, code, databases, tools, benchmarks, and algorithms, then packages them as executable skills with declared dependencies, interfaces, inputs, outputs, and execution steps. Skills are syntax- and dependency-checked and validated through isolated usage examples before reuse. The resulting library supports domain-specific molecular property prediction, protein folding, structure prediction, sequence design, scoring, and data processing.
What the paper evaluates
For antibiotic prediction, MolRSI was evaluated on six tasks covering S. aureus, P. aeruginosa, A. baumannii, E. coli, K. pneumoniae, and N. gonorrhoeae. Twenty-five source tasks—22 Therapeutics Data Commons tasks and three cytotoxicity tasks for HepG2, HSkMC, and IMR-90—provided experience, while the six antibacterial targets were held out from that memory pool. All methods used the same five similarity-based splits; mean validation AUPRC selected workflows, test AUPRC was reported across splits, and each task allowed up to 20 workflow evaluations. MolRSI without RSI achieved the highest mean test AUPRC, 0.361, and ranked first on four tasks and second on two.
Limit to keep in view
The prospective antibiotic experiment uses one screening library and a specific MSSA hit threshold, so generalization to other compound libraries, organisms, and experimental protocols remains to be established.
arXiv:2611.08318
Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement
This paper studies whether exposing an embodied self-improving agent to benchmark success criteria improves task learning or mainly teaches it to satisfy a verifier. In a controlled RoboDojo study, a fixed coding actor and learner repeatedly build skill-and-memory libraries with either success-only feedback or success plus criteria code, scores, and stages. Under original criteria, Criteria Exposure raises overall test success from 47.8% to 62.0%, with the gain concentrated on criteria-limited tasks (51.6% versus 27.8%) rather than faithful tasks (80.6% versus 83.3%). Rejudging the same episodes with instruction-faithful criteria removes the advantage (45.0% versus 47.2%), and exposed libraries provide no clear transfer advantage across new tasks or benchmarks.
How it works
A task is represented as an instruction, layout distribution, and success criteria. The actor is frozen GPT-6 Astra with high reasoning effort, and the learner is Claude Code with Opus 5.5 and high reasoning effort. Each task has its own library of callable skills and memory entries, learned over five rounds on five development layouts. In w/o CE, the learner sees only whether each episode succeeded; in w/ CE, it also reads criteria code, scores, and stages.
What the paper evaluates
The main study learns one library per task on 25 RoboDojo tasks: 16 criteria-limited and 9 faithful. Development uses layouts 0–4 of seed 1, testing uses layouts 0–19 of seed 0, and the metric is success rate under the relevant criteria. Table 1 reports the original-criteria comparison: success-only learning improves all tasks from 24.6% to 47.8%, while exposure reaches 62.0%; on criteria-limited tasks the comparison is 27.8% versus 51.6%, whereas on faithful tasks it is 83.3% versus 80.6%. The development curves in Figure 3 show both settings starting at 6% on criteria-limited tasks, separating in round 1 at 40% w/ CE versus 28% w/o CE, and ending at 55% versus 31%.
Limit to keep in view
The instruction-faithful criteria are manually authored by the study and act only as a proxy for the success condition implied by the instruction; only RoboDojo criteria were rewritten.
arXiv:2611.06321
EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design
EnzymeRSI proposes a two-loop system for enzyme engineering: an outer loop revises prompts, pipelines, tools, and other workflow assets while keeping model weights fixed, and an inner loop develops prediction or selection methods when the method itself is the bottleneck. Across five ProteinGym assays, the outer loop increased mean development-set Spearman correlation by 0.044; only its CD19 result produced a first-place comparison, while OPSD’s first place was already present at the starting point. On ThermoMPNN, the inner loop reached a development-set PCC of 0.849; on TadA-Bench it reached Spearman 0.307 in 1.13 hours. Three prospective enzyme studies tested nomination, constrained selection, and evolutionary path choice.
How it works
The task contract specifies the parent sequence, objective and non-negotiable constraints, experimental context, legitimately accessible data, and multidimensional budget. The policy selects actions given the research state and the active harness and algorithm versions. Trajectories record the evidence available at decision time, hypotheses, actions, results, and subsequent decisions. The outer loop diagnoses workflow failures and proposes a minimal, falsifiable change bundle.
What the paper evaluates
**Computational evaluation.** The inner-loop development tasks were ThermoMPNN stability prediction and the TadA-Bench amino-acid view. ThermoMPNN used 215,731 training records from 239 proteins and 27,328 development records from 31 proteins; the official predictor was the common starting point, with development PCC 0.804358. Each run had 20 formal evaluations, a 48-hour research wall-clock limit, and a 96 GiB memory limit. Against ten reference frameworks, EnzymeRSI reached PCC 0.849, above the highest reference, OpenRSI at 0.834; the paper associates the gain with introducing ESM2 representations.
Limit to keep in view
The outer-loop ProteinGym gains are development-set results; only the CD19 first place was produced by evolution, while OPSD already led at initialization.
Read the full reviews
- Learning Multi-Agent Orchestration:Higher-Order Interactions in Policy Improvement · arXiv:2611.07495
- MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery · arXiv:2610.39714
- Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement · arXiv:2611.08318
- EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design · arXiv:2611.06321
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.