Zixun Huang, Kishan Panaganti, Haitao Mi +1cs.LG cs.AI
A reasoning model can improve from its own on-policy experience, but this inner loop is fragile: terminal verifiers provide reliable yet sparse supervision, while dense same-model guidance can reinforce false confidence or overconcentrate learning on a narrow solution mode. We introduce FlowBalance, a verifier-grounded self-improvement method that learns a normalized distribution over complete responses. For each on-policy trajectory, a frozen training-time view of the same policy uses privileged context to produce token-level log-probability gains, which are aggregated into a trajectory-level self-guidance score. FlowBalance calibrates this score with the verifier-derived group advantage: guidance is retained on positive-advantage trajectories, reversed on negative-advantage trajectories, and disabled when the rollout group provides no outcome preference. The resulting energy exponentially reweights a reference policy, and profiled trajectory balance fits the normalized target with one log-partition estimate per rollout group. This realizes outcome-calibrated self-guidance via trajectory balance, without a separate token-level imitation loss. Our analysis establishes within-group contrast preservation, a minimum-change reverse-KL characterization, monotonic verifier control of target reward, and an exact correction against false-positive self-guidance on rejected responses. On mathematical reasoning, FlowBalance improves average performance over FlowRL on both Qwen3-4B and Qwen3-8B, while also improving training speed and stability, avoiding direct OPSD's response-length collapse, and exhibiting higher correct-strategy diversity in a controlled AIME24 diagnostic.
Huaiyuan Yao, Xiaoou Liu, Charles Fleming +2cs.AI cs.CL
LLM-based multi-agent systems have shown strong performance on complex tasks, yet continual improvement from interaction experience remains challenging. Existing self-reflection methods build experience memories, but memories are mostly hard to invoke, refine, or scale, while agent skills offer a more actionable unit: structured procedural knowledge that specifies when to act, how to act, and which resources or tools to use. We introduce MASkills, a continual learning framework that optimizes multi-agent LLM systems through agent skills. MASkills presents a new agent-optimization pipeline that integrates skill-conditioned credit assignment, hierarchical credit aggregation, and momentum-smoothed optimization, enabling agent skill libraries to evolve through refinement, induction, consolidation, and pruning. Experiments on HotpotQA, LoCoMo, and GAIA demonstrate the effectiveness of MASkills across multiple agentic tasks. Our code is available at https://github.com/DaRL-GenAI/MASkills
Yuhao Wu, Jingyuan Zhang, Jiajun Shi +16cs.SE cs.CL
As agents move from research prototypes to deployed tools, their capability increasingly depends on model-external execution infrastructure, commonly termed the agent harness. Changing this harness while holding model weights fixed can substantially alter task performance. Current agent evaluations typically report downstream performance under a chosen harness, leaving a model's ability to develop the harness itself comparatively underexplored. We introduce HarnessDev, a benchmark that shifts the unit of evaluation from task outputs to runnable infrastructure. HarnessDev covers two stages. In Creation, the agent starts from a minimal seed and a small number of cases, then builds a complete execution system. In Evolution, it starts from its own created harness and iteratively revises it using downstream execution feedback, with the goal of improving benchmark performance. We then evaluate each constructed harness on capability (task success on held-out benchmarks) and efficiency (execution-token cost). The reported Creation results cover six creator LLMs, four domains, and five downstream benchmarks totaling 2,207 unique downstream instances, with hidden evaluation tasks withheld from development. We find that generated harnesses remain substantially behind mature human-engineered references on code and on search and research, while matching or exceeding the selected references on writing and machine-learning experimentation, with large variation in execution cost. Evolution produces some performance gains, but they are unstable and transfer only partially to held-out tasks. Experiments with a fixed runtime model further show that the gains depend strongly on the model executing the harness, indicating limited transfer across models.
Inference cascades cut cost by answering most queries with a cheap model and escalating a hard tail to a frontier model that acts as verifier. A natural extension closes the loop: fine-tune the cheap student on the verifier's rejections so the escalation rate, and cost, fall each round. We measure this loop on real LLMs and report four findings. First, the verifier's blind spot, the fraction of the student's wrong answers it accepts, is large and moves adversarially: it grows with student capability ($β$ from 0.12 to 0.55 as the student scales 0.5B to 32B) and shrinks with verifier capability, so it is worst in the cheap-student, cheap-verifier regime cascades exist to create. Second, buying it away returns the saving: a frontier verifier drives $β$ to about 0.05 but then escalates on 46% of hard-MATH queries against a 39% true error rate, paying the frontier price on nearly half of all traffic. Third, naive corrective fine-tuning on the verifier-rejected tail does not improve the small student but degrades and ultimately collapses it, across every teacher we tried (cross-family and same-family), so at this scale the self-improving loop is self-defeating. Fourth, through all of this the cascade's own dashboard, every metric computed through the verifier, reads a flat 3% error while true delivered error swings up to 32%: the system is blind to its own degradation by construction. We then give the theory that explains the blindness, a two-population conservation law, $ε_\infty \lesssim q_0 β_0$, under which every in-loop metric improves while true quality does not, and a synthetic study that validates the mechanism. The practical conclusion: the reliability of a self-improving cascade cannot be read from any metric computed through its own verifier.
The ability of AI systems to improve their behavior during deployment is becoming increasingly important. As inference moves beyond the static execution of a fixed trained model, a growing body of work studies how models can refine their behavior on the fly by exploiting test-time information and additional computation. These developments have largely evolved along two directions: methods that modify the model's state using test-time signals, and methods that improve predictions through extra inference-time resources such as more sampling and tool use. However, these directions are often studied in separate communities with different terminology, making their connections harder to see. In this survey, we present feedback-driven Test-Time Intelligence (TTI) as a unified perspective for understanding such deployment-time improvement. We use this view to relate test-time adaptation, test-time learning, and test-time scaling, highlighting both their distinctions and their growing overlap in hybrid systems. This unified framework helps connect previously fragmented ideas and provides a clearer conceptual foundation for studying inference-time self-improvement. We review major methodological paradigms, representative applications, and open challenges across vision, language, multimodal learning, generative models, robotics, and healthcare. Our goal is to provide a coherent foundation and research roadmap for the study of self-improving AI systems at test time.
Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. We argue that an ideal self-evolution method should share the same property, that is autonomously learning from raw training material for transferable problem-solving capability. However, we still lack a direct measurement for it. We introduce StudyBench, a controlled physics benchmark that directly measures how efficiently a self-evolution method converts training material into capability. We organise the test set into an Application Set, consisting of difficult textbook problems and evaluating absorption ability, and a Transfer Set, consisting of olympiad-level problems and evaluating transfer ability. Benchmarking representative self-evolution methods across three base models, we find that improvements on the Application Set rarely translate to the harder Transfer Set. A guidance ablation exposes a Guidance Gap: even the strongest method closes only a small fraction of what the same material unlocks when supplied as in-context guidance. Besides, every method hits a Compute Plateau, saturating well before exhausting its compute budget. The remaining gap is therefore a method problem rather than a data or compute problem. By offering a clean and controlled benchmark, StudyBench turns self-evolution progress from an open-ended pursuit into a measurable target for future research. Our code is released at https://github.com/thunlp/StudyBench.
Large language models (LLMs) increasingly interact with external environments and accumulate substantial behavioral experience, yet existing agent benchmarks largely evaluate them as fixed policies. It therefore remains unclear whether an agent can actively test its behavior, judge the resulting experience, and use that experience to improve future decisions. We introduce \textbf{S\textsuperscript{3}Gym}, an interactive benchmark for evaluating LLM self-improvement through three coupled capabilities: \textbf{Self-Testing}, \textbf{Self-Judging}, and \textbf{Self-Improvement}. S$^3$Gym separates permissive exploration from strict held-out evaluation and instantiates this protocol in seven text-based games with executable environment verifiers. We evaluate three pathways for incorporating interaction experience: direct History ICL, score-conditioned Summary Memory, and parameter Training. Our experiments reveal that self-improvement is neither automatic nor uniform. Context-level experience improves performance for several model--game pairs, but the most effective pathway depends strongly on the task structure: summaries are beneficial when experience can be compressed into reusable strategic rules, yet often underperform raw history when success depends on precise, state-contingent information. Parameter training produces substantial gains on some tasks, but also exhibits unstable improvement and severe negative transfer on others. These findings show that recognizing successful actions is insufficient; agents must also transform feedback into executable and transferable policies. S$^3$Gym provides a unified framework for diagnosing this process and identifying the bottlenecks that prevent agents from translating interaction experience into reliable self-improvement.
Recent advances in large reasoning models (LRMs) have shown that reinforcement learning with verifiable rewards (RLVR) can substantially improve reasoning in mathematics and code, where outcomes can be checked automatically. Extending this progress to open-ended and agentic tasks remains difficult because reliable rewards are harder to obtain and direct human supervision cannot keep pace with the scale and complexity of model-generated experience. This paper studies how LRMs can continue to improve as human supervision gradually recedes from the learning loop. We examine two connected dimensions of this problem. The reward axis traces the development from per-instance human judgments to reusable verifiers and rewards that operate even without human feedback. The experience axis examines how learning can progress from human-curated tasks and environments toward self-generated curricula, constructed environments, and autonomous co-evolution. We connect these dimensions through a five-level ladder from L0 to L4 that identifies which parts of the learning process remain under continued human control. Our analysis further highlights the risks introduced by increasingly autonomous rewards and experience generation, including reward hacking, feedback drift, curriculum collapse, and environment errors. Consequently, we also provide the evaluation around three complementary objects: policy capability, feedback fidelity, and experience quality. This analysis provides a structured account of current approaches to scaling LRMs beyond human supervision and the open problems involved in developing self-sustaining learning systems toward superintelligence. Furthermore, we maintain a continuously updated \href{https://github.com/visitworld123/Awesome-Scaling-LRM-Beyond-Human-Supervision}{GitHub repository} to track the latest advances.
VLM-driven self-improvement of web code has a structural flaw: the model that proposes the repair is the model that judges it, and visual plausibility under that judge is a poor proxy for whether the page actually works. What the loop is missing is a counterparty the VLM cannot fool, and the browser already is that counterparty: a deterministic, executable simulator of how an HTML artifact behaves under user actions, and in everything but name a world model for web code. We present WebWorld, the interface that lets a VLM prior interact with this browser-as-world-model autonomously and decides which interactions become supervision. Each round, the VLM emits a critique that the planner compiles into a typed interaction contract; the browser re-executes the candidate and issues an acceptance certificate only when both target progress and preservation of every previously verified capability hold; certified transitions accumulate as a quality ratchet that is the only thing the SFT export ever sees. Under matched training, WebWorld-27B improves Raw-27B by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val, and reaches the level of strong frontier systems such as Kimi-K2.6 and GPT-5.4 on interactive HTML generation. Equal-size ablations show that browser-backed admission carries the gain: without the certificate, the matched 9B lift nearly disappears.
Trajectory data is getting more vital for training large language models for boosting the agentic abilities. Unlike the verifiable domains such as coding or mathematics, scaling trajectory data for open-ended tasks is much more difficult because these tasks lack singular ground truth and are costly to annotate or verify. In this paper, we propose RetroGen, a self-improving framework of retrospective process supervision. Our key observation is that although expert trajectories are scarce, high-quality final artifacts such as literature reviews, analyst reports and legal judgments, are abundant in pre-training data and can be viewed as compressed traces of the evidence-seeking processes that produced them. RetroGen reconstructs candidate latent trajectories from expert artifacts, verifies them against both the artifact and supporting evidence, and trains models on their own successful reconstruction data, without requiring trajectory data from stronger models. Experiments show that RetroGen improves grounding, faithful synthesis, and long-form evidence-seeking agent tasks.
Current dialogue systems struggle with dynamic information retrieval, often leading to hallucinations and lower response accuracy. We address this by adapting the ReAct framework for Task-Oriented Dialogue, enabling Large Language Models (LLMs) to access external knowledge and produce factual responses. Mainly, we propose an unsupervised fine-tuning pipeline that harvests reasoning trajectories via in-context learning inference. High-quality samples are filtered using an LLM-based judge to construct a robust training set. This is enhanced by a unsupervised self-improvement loop, where improved checkpoints generate increasingly better trajectories for subsequent fine-tuning iterations. Experiments on the SIMMC dataset demonstrate that ReAct-based systems outperform baselines due to superior reasoning and tool use. Notably, our fine-tuned 8B model surpasses a 70B in-context system. Finally, we present an error analysis, impact of scene complexity, and cross-domain generalization.
Domain adaptation of large language models increasingly depends on constructing high-quality training data, yet existing data-preparation pipelines typically address quality only after generation through post-hoc filtering. This creates a fundamental mismatch: data-quality issues often originate from the construction process itself, while quality control is applied only to its outputs. We introduce \textsc{DataFoundry}, a framework for \textbf{evolving data preparators through recursive self-improvement} before large-scale data production. \textsc{DataFoundry} represents a data preparator as an evolvable runtime specification and instantiates its evolution with a \textsc{Skills-as-Modules} architecture, in which a central \textsc{Controller} orchestrates modular skills to compile executable runtimes, diagnose deficiencies on small pilot sets using domain-appropriate criteria, and translate diagnostic feedback into adapters that revise individual preparation components while preserving stable interfaces. We evaluate \textsc{DataFoundry} on DataPrep-Bench across mathematics, finance, law, and medicine, and find that recursively evolved preparators produce training data with higher downstream utility than baselines. Experiments across different backbones further demonstrate that these improvements are not tied to a particular model, while analyses and case studies further reveal the framework's optimization dynamics and illustrate how its evolution unfolds in practice.
Long-horizon story-driven video generation requires a production agent to coordinate narrative decomposition, state tracking, shot design, prompt construction, rendering, and revision across interdependent scenes. Existing adaptive video systems primarily refine requests or reusable skills, leaving recurring production failures disconnected from persistent, stage-targeted improvements across stories. We introduce CineForge, a self-evolving video-production agent framework that couples CineForge-Produce for video generation with CineForge-Evolve for cross-story policy evolution. CineForge-Produce organizes each source story into typed narrative, character, spatial, and cinematic states, uses them to coordinate asset and clip generation, and records the process as a canonical production trajectory. CineForge-Evolve applies Case-to-Pattern-to-Policy Evolution (CPPE) to review trajectory evidence, consolidate recurrent findings into bounded stage-local patches, and deploy validated updates through structural replay and confidence-controlled paired evaluation. To measure complete story realization, we introduce CineScope, which combines a 100-script CineScope-Data suite with a human-aligned, multiscale CineScope-Metric spanning causal state, directorial orchestration, pacing and resource allocation, and character arc. Across CineScope-Data and two public benchmarks, the evolved CineForge policy improves CineScope-Metric from 4.024 to 4.380, outperforms three long-video baselines with consistent gains under ScriptAgent, and reduces review LLM calls by 37.0% on new stories. These results establish production trajectories as actionable experience for video agents that improve cumulatively across long-form storytelling tasks.
What does a generation loop gain from learning on its own verified successes? In cycles of generate, verify, select and LoRA-consolidate on online bin packing, training on value-filtered candidates shifts what the model writes on held-out variants toward value (-1.7 points of excess, p=0.008; -3.1 against a random-consolidation control, p=0.004) while the best observed candidate converges to the classic heuristic's level and no further. A confirmation battery replicates the whole procedure three times, with fresh seeds and a never-consulted held-out set read exactly once: the mean was nearly identical in all three lineages (-2.0, -1.8, -1.9), and after aggregating within held-out variant all seven evaluable variants favored consolidation (p=0.008). The best observed candidate moved to the classic heuristic's level, exactly (0.021028 in all three lineages, for attract and for the random control alike), and never beyond it. A matched SFT-only control shows the supervised anchor, not repulsion from bad candidates, does the concentrating (96% of candidates land exactly at the classic heuristic's level). The tails cut both ways: consolidation lowers the per-candidate rate of better-than-classic candidates (10% to 3.9%) while its larger production yields more such candidates absolutely (5 against 1, on few events). As motivation we report the inference-time ledger that led here: a model-written schematic recap buys judged document integration and nothing buys development; a verifier written into the stream is imitated, 16.4 fabricated verdict lines per notebook. Mean quality among valid candidates can be bought and replicated; the observed best goes to the classic and, so far, never beyond it.
Achieving virtual agents capable of automating tasks across diverse digital environments remains a pivotal challenge in Embodied AI. While Multimodal Large Language Models (MLLMs) offer enhanced visual perception and reasoning, their agentic deployment faces three challenges: costly data annotation, imprecise action-intent alignment, and inefficient exploration from discarded failed trajectories. To address these, we introduce Iron, an intent-aligned, self-improved, and annotation-efficient framework for training GUI agents. Iron employs a novel dual learning strategy that utilizes a stepwise cycle-consistent (SCC) reward to achieve fine-grained alignment between low-level actions and high-level intents, thereby improving instruction grounding and intent understanding. Concurrently, Iron introduces a hindsight reproduction mechanism to repurpose failed trajectories for training, improving both learning efficiency and task diversity. Extensive experiments demonstrate that Iron-trained generalist agents consistently improve performance on cross-environment and cross-device tasks, outperforming models trained with three times more data. Iron also achieves a substantial 25.06% relative improvement on unseen web tasks, with further gains observed on inherently complex tasks, demonstrating the feasibility of building more capable virtual agents.
Long-horizon agent runs generate experience that can improve both the current run and future work. Most self-improvement methods process this experience only after execution ends, so they cannot redirect the active run or immediately apply and validate lessons learned from it. We argue that self-improvement should instead be live, using emerging experience both to redirect the active run and to update the persistent harness. Existing agent architectures do not fully support this goal. Single-agent self-correction combines task execution and trajectory assessment within one context, while subagent delegation separates execution but typically cannot redirect an active subagent. We present PILOT, a supervisor-worker harness for live self-improvement through two coupled mechanisms: (1) live steering lets a separate supervisor redirect or abort the active worker during execution; and (2) live self-evolution distils procedures and failure modes revealed during execution into reusable skills and memory. Across two frozen backbones and three benchmarks, PILOT ranks first in five of six configurations. On Terminal-Bench 2.0, PILOT outperforms counterpart harnesses by up to 9.8 percentage points. In the self-improvement setting, PILOT gains 14.6 points with GLM-5.1 and 12.4 points with Kimi-K2.6. Mean output tokens fall by 42.9% and 47.4%, while successful evaluations per million output tokens rise by 110.3% and 134.0%, respectively.
Outcome-supervised search agents learn when and how to retrieve evidence, but terminal rewards neither localize intermediate errors nor redirect an ongoing trajectory before those errors compound. Treating corrective feedback as a learned in-trajectory intervention couples the two roles: the agent must decide when to request and use feedback, while the critic must infer useful corrections from outcome-confounded rollouts whose failure patterns shift as the agent improves. We introduce CAFE (Coupled Agent--Feedback Evolution), a framework in which a shared-parameter model alternates between search-agent and critic roles. CAFE initializes feedback-conditioned recovery from trajectories built around the base agent's own failures, then couples online and offline optimization. During online RL, a comparative feedback estimate uses a prompt-level call--skip success gap to shape request returns, while feedback-aware advantage shaping reweights token advantages before and after feedback. Offline, rollout-derived preference optimization learns feedback from matched successful and unsuccessful trajectories. On seven agentic search benchmarks, CAFE outperforms the evaluated RL-based search agents on average, retains its gains across all six out-of-domain benchmarks, and reduces answer-level hallucinations. One-sided ablations show that improving only the agent or only the critic eventually plateaus, whereas alternating the two updates continues to improve performance. These findings suggest that a self-improving search agent needs feedback that co-evolves with the policy it guides.
Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa +1cs.AI cs.CL eess.SY
Self-improving LLM agents refine answers, not the process that produces those answers. Systems that add a meta-level hold that level fixed, and those that edit themselves must leave part of their own editing machinery untouched to stay stable, capping the meta-depth they realize at roughly two. We present Meta$^n$, which keeps the meta-operation fixed and recurses on its input instead. That operation, $Ω$, is applied repeatedly to its own products, reading the traces of the solver stack below together with the code that produced them, then writing the next layer as a strategic pre-process and a library of callable helpers. Because $Ω$ never changes, it cannot destabilize the system, and because its input strictly grows, each layer reasons from a higher vantage than the last. Depth is set by convergence rather than fixed in advance, and an evolutionary archive searches over layer chains. Across two backbones, Meta$^n$ outperforms prior self-improving agents on all eight benchmark families. The sharpest case is ARC-AGI-2, built to resist skill memorization, where it alone scores above zero. Ablations indicate that most of the gain from recursion comes from the conditioning each layer passes to the next, and distinct layer roles emerge with depth although no prompt prescribes them. Code available at https://github.com/minnesotanlp/meta-n
Long-video understanding depends critically on how a limited model context is constructed from a much longer video. Existing approaches improve this process through compression, retrieval, memory, and agentic evidence acquisition, but these mechanisms are typically introduced as part of a manually designed inference system or optimized together with other components. This makes it difficult to isolate a simpler question: how much can be gained by improving the executable context-construction program alone? We study this question through VIDEOHARNESS-RSI, a controlled baseline for recursively searching executable context constructors around a frozen vision-language model (VLM). An outer-loop proposer uses prior programs, evaluation outcomes, and execution traces to generate candidate harnesses, which are executed and evaluated end to end before successful variants are retained for further search. This makes long-video understanding a controlled instance of automated harness design: the searchable object is executable program structure, while the answering model and interface remain fixed. Starting from uniform sampling, recursive harness search consistently finds room for improvement and surpasses several weaker hand-crafted baselines. Starting instead from a stronger hand-crafted baseline, the same RSI process yields a further improvement. The selected harness also transfers to additional long-video benchmarks without further search. Together, these results establish executable context construction as a distinct optimization layer and provide a reproducible baseline for studying harness discovery and transfer around frozen VLMs.
LLM-as-judge is essential for evaluating open-ended text and steering post-training, yet improving the judge itself typically relies on expensive annotations, reward models, or distillation from stronger teachers. In this work, we eliminate external gold supervision from the RL training reward: the model's own evaluative capability generates learning signals for its optimization -- a closed-loop setting of bounded recursive self-improvement (RSI) termed Recursive Self-Evaluation (RecurSE). We study two central questions: when can self-improvement occur, and when must it stop? First, RecurSE pairs a trainable judge evaluating candidate responses under per-rule rubrics (Pass 1) with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward (Pass 2). To enable learning, interface decoupling structurally isolates the checker's scalar score from the judge's verdict tokens, eliminating a degenerative token-copying shortcut that inflates self-assigned rewards. Second, because unanchored recursive learning is inherently bounded, Pairwise Advantage Validity (PAV) serves as an unbiased validation monitor that jointly tracks judge accuracy and checker fidelity to reliably identify the optimal early-stopping window. Across Qwen3.5-9B, Gemma-4-E4B-it, and Qwen3.6-27B, RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks. Ablations demonstrate that synchronized judge-checker co-evolution outperforms frozen checkers, external meta-judges, self-consistency, and scaled teacher distillation. Furthermore, preference pairs curated by our judge effectively enhance downstream policy alignment. Bounded RSI for LLM-as-judge is thus viable when self-produced reward validity is explicitly decoupled and monitored.
Seth Karten, Alex L. Zhang, Kevin Thomas +8cs.AI cs.CL cs.SE
Language models are sequential processors, but long-horizon agency requires external information and computation beyond model weights and active context. Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows. A persistent IPython REPL follows the Recursive Language Model abstraction for programmatic context processing and test-time compute, while Continual Harness preserves histories, memories, skills, prompts, and subagent specifications across trajectories. Recursive subagents coordinate through direct agent-to-agent communication, and the Agents View lets humans inspect and manage daemon-backed sessions. Prime Agent standardizes execution, recovery, verification, and resource accounting while leaving strategy construction to the model. This low-friction, expressive membrane prevents harness failures from becoming model failures and pushes measurement toward the model's true maximal underlying capability. Prime Agent raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5% and matches or exceeds native and popular harnesses across long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns. On Factorio, we find refinement allows for continuous technology progression and dedicated subagents enable parallelized work. Code is available at https://github.com/PrimeIntellect-ai/prime-agent.
Joe Yu, Shibin Thomas Stanley Paul, Sven Mayercs.AI
Automated prompt and skill optimization typically produces a single static instruction that is reused across inference instances until the next optimization cycle. However, this approach cannot adapt when the required context, constraints, and evidence vary from one instance to another. For instance, parameter extraction from electronic component descriptions breaks this assumption: resistors, capacitors, transistors, and connectors require different fields, unit constraints, and demonstrations, and each input provides a different evidence state. We introduce DynaContext, a framework that combines an offline-optimized extraction core, learned with GEPA or SkillOpt, with inference-time contextual adaptation and validation-gated self-improvement. DynaContext routes each item through internal, external, or fallback evidence paths and composes an item-specific prompt from the core, schema, evidence, unresolved fields, and validated demonstrations. Deterministic validation and an LLM judge gate every output, uncertain cases go to human review, and only human-verified corrections enter the demonstration memory. On a single-category benchmark, average accuracy increases from 86.6% for the base prompt to 96.9% for standalone SkillOpt and 98.6% for the best DynaContext configuration. Across 850 heterogeneous gold parameter facts, average field-level F1 increases from 51.8% for an unoptimized, demonstration-free control to 59.2% with dynamic demonstrations alone, 66.9% with the optimized core alone, and 71.0% with both. Holding the model fixed, the full configuration outperforms the deployed static-prompting pipeline by 17.3 F1 points on average.
Varun Giridhar, Anant Khandelwal, Jeremy A. Collins +2cs.RO cs.LG
Behaviour Cloning (BC) has driven remarkable progress in robot manipulation, yet it is fundamentally limited by its inability to self-improve: a policy that fails cannot learn from that failure without additional human demonstrations. Reinforcement Learning fine-tuning offers a path to self-improvement but has proven difficult to scale to the multi-billion-parameter models underpinning modern robot policies. We propose Q-Planning, which equips a large visuomotor BC policy with a small off-policy Q-function. Because a Q-function estimates value rather than imitates actions, it can be trained on the same successful demonstrations as the BC policy and later absorb both successful and failed deployment rollouts, an asymmetry BC does not have. We exploit this asymmetry to enable value-guided action selection at inference (a single-step Q-weighted average over BC draws) and online self-improvement that fine-tunes only the Q-function, leaving the BC weights untouched. On LIBERO and bimanual RoboTwin, ten iterations of self-improvement lift every benchmark score we tested (LIBERO-10 93% to 99%, RoboTwin 83.8% to 91.4%) and shorten successful episodes on the near-ceiling suites (LIBERO-Object, LIBERO-Goal). On two contact-rich bimanual real-robot tasks, the same loop (BC frozen, no human intervention) improves purely from its own deployment rollouts: stack-cups 40% to 90% and insert-wallet 25% to 80% in five iterations, whereas SFT on successful rollouts alone stalls at 55% and 30%. Under an identical online budget Q-Planning is the only method, among Best-of-N, filtered SFT, IBRL, DSRL, and DAWR, that improves stably from failures without training an auxiliary actor.
Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-$32$ LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of $0.280$. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain ($p < 10^{-8}$). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.
Continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals. For language agents, existing training environment pools (hand-curated, statically synthesized, or frozen-verifier) keep the goal distribution fixed as the learner scales. We introduce SPADE (Self-Play in Adaptive Synthetic Executable Environments), a self-play RL framework in which a single LLM plays two roles: an Environment Designer that writes complete, long-horizon training environments as executable code with an OpenAI Gym-style reset()/step() interface, and a Reasoning Agent that learns to act in them. Each is a stateful, multi-turn environment (state transitions, reward functions, and verification code), so one interface spans reasoning problems and multi-step agentic tool use. The Reasoning Agent's regret is estimated using the gap between its reward with and without privileged hints; in optimizing this regret signal the Environment Designer learns to target environments at the edge of the agent's capabilities while keeping them feasible. Through extensive experimentation, we find several components critical to success: grounding the Environment Designer on documents sampled from a large pretraining corpus, and giving it an accumulated environment memory. Scaling to 30B-parameter models, SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks, and lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent; on the games setting, the margin over the strongest baseline grows with model scale. By making environment design itself a learnable component, SPADE takes a concrete step toward open-ended self-improvement.
Long-horizon agents are beginning to automate complete workflows that produce code, reports, and research artifacts. Medical imaging workflows are multi-stage and data-sensitive, while expert trajectories remain scarce and difficult to share. Structured benchmarks can localize failures through stage-level rubrics, but standard post-training discards these diagnostics before the next training round. We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: the asynchronous Stage Bank data pipeline and BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method. Stage Bank synthesizes content-isolated training states outside the policy-update loop. BiCuRL uses a fixed held-out evaluation to select the next stage curriculum, verifies rollouts with task rubrics, updates the policy with GRPO, and returns the candidate checkpoint to evaluation. On AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines. BaT-9B Agent reaches 79.6 Overall, exceeding Claude Opus 4.6 with Claude Code at 77.5.
Scaling agent capability has largely focused on improving the model, yet an interactive agent acts through a runtime harness that mediates context, tools, control flow, and stopping. The harness shapes both what a model can accomplish and the trajectories from which it learns. This coupling motivates model-harness co-evolution for recursive self-improvement: build harnesses for a fixed model, update the model from verified sibling trajectories, and rebuild the harnesses as model capabilities change. Realizing this loop requires a controlled way to evolve harnesses while preserving intervention identity and effect. We present HELIX, a source-traceable substrate for harness evolution. HELIX decomposes agent systems into typed ports, reusable atoms, recipes, product shells, and runtime policies. It makes interventions explicit and auditable while retaining trajectories, test outcomes, and provenance. Harness evolution thus serves two linked roles: improving fixed-model execution and producing matched successes, regressions, near misses, and alternative solutions as data for subsequent model improvement. We evaluate HELIX in one evolution round on code repair. A 65-candidate portfolio discovers a fixed harness that improves task coverage by 4.0% over Pi, while the full portfolio exposes up to 58.0% more verified coverage through complementary sibling behavior. Selected candidates are assessed with repeated runs and the SWE-bench evaluator. A 200-slot sibling slice yields 438 verified SFT, critic, filter, and preference records. These results show how harness, model, and data form a feedback system: harness evolution expands current capability and creates learning signal for the next model; model updates motivate the next round of harness evolution. HELIX provides an auditable interface for studying this recursive process. Code is available at https://github.com/HKUDS/HELIX.
Agent Skills are today either hand-authored or produced in a single LLM generation pass, and consequently possess no closed loop through which they might improve from the interaction failures they actually cause. Recent work does close this loop, but derives its feedback from single-turn question-answering evaluation. The consequence is a sharp asymmetry: once the first round has patched the gaps that a single exchange can reveal, the evolution gradient decays, the defects that surface only across multiple turns remain invisible, and evolution stalls. Governance in these systems is likewise driven by an end-to-end verification score, a scalar gate that can reject a degraded candidate but can neither localize nor repair its structural cause. We argue that the binding constraint on sustained skill evolution is neither editing capability nor the number of iterations, but whether the evaluation feedback keeps supplying trustworthy evolution gradients. We introduce SkillEvo, in which trustworthy feedback generates the gradient and controllable governance constrains its direction. The first component recasts multi-turn user simulation from an evaluation endpoint into a feedback generator: follow-up questions expose defects layer by layer, so that every round of revision both consumes feedback and produces new feedback. The second replaces the passive rejection of a scalar gate with an independent governance layer that actively repairs factual degradation and structural bloat, preventing the gradient from drifting as degradation accumulates. Across six categories of cloud services, 9 production Skills, and 98 skill-reference files, SkillEvo surpasses self-reflection-based evolution by 23.0 points and single- turn-QA-driven evolution by 15.4 points.
Self-improving LLM agents convert successful trajectories into persistent cross-task state. An unsafe success can thereby become reusable policy after its triggering input disappears. Skill evolution makes this failure measurable by distilling operational trajectories into executable, transferable, and inspectable procedures. Because evolution optimizes task outcomes rather than procedure safety, compromised experience can cause skill misevolution. Existing benchmarks measure current behavior or static artifacts but cannot attribute risk across authoring, retrieval, and later execution. To expose this lifecycle, we introduce SkillMisevo-Gym, a lifecycle-aware harness that versions skill state across agent frameworks, and SkillMisevo-Bench, a frozen design from malicious exposure to carryover tasks, with concept-aligned benign tasks and nine lifecycle metrics. We also introduce SafeEvolve, a wrapper that repairs unsafe content and governs subsequent reuse. Across 25 agent-method configurations, each covering 525 tasks in 25 episodes, all 21 evolved configurations author unsafe artifacts, while only fifteen lead to fresh-session harm. In the exposure sweep, three malicious tasks raise carryover ASR from 16.0% to 35.3%. Across representative skill evolution methods, SafeEvolve reduces unsafe retrieval and fresh-session harm by 26.7 and 17.3 percentage points, respectively, while mean benign utility changes by only 0.4 points. Together, persistent-adaptation safety must govern what updates write and what future executors reuse. Code is available at https://github.com/henrymao2004/misevolve.
Siheng Xiong, Ali Payani, Oguzhan Gungordu +1cs.CL
Large language models (LLMs) cannot retain post-deployment experience without parameter updates. We introduce DIVE, a diversity-driven framework that enables frozen LLMs to improve by evolving persistent natural-language skills from task experience and verifier feedback. These skills encode reusable reasoning procedures, verification strategies, common failure modes, and output constraints and are both executed and revised by the same underlying model without access to a teacher model. Since natural-language skill evolution is a stochastic, non-convex search process, optimizing a single skill trajectory can overfit to sampled experience or converge to a suboptimal solution. DIVE mitigates this optimization variance by independently evolving multiple skill populations from bootstrapped experience, adaptively refining them through diverse transformations, and jointly selecting a complementary set of skills. Across six mathematical and logical reasoning tasks and multiple model families, DIVE consistently outperforms existing reasoning methods, prompt-optimization approaches, skill-development frameworks, and memory-based baselines. It achieves rapid self-improvement from accumulated experience, obtaining substantially larger performance gains with fewer rollouts than parameter-based methods such as SFT and GRPO, and prompt optimization with GEPA. Further, the resulting skills transfer across model scales and families, enabling smaller models such as GPT-5-nano to match or outperform larger counterparts, i.e., GPT-5, under conventional prompting. These results establish diversity-driven skill evolution as an effective, interpretable, and parameter-free approach to LLM self-improvement.