Evolution

Data agents aim to fulfill natural-language instructions over heterogeneous data, including tables, files, and databases. However, data agents face a challenging agent-data gap: heterogeneous data resides outside the agent, while the agent can access it (e.g., column names and file paths) only through generic tools. Existing approaches either let agents directly explore raw data sources or inject manually constructed semantic layers into prompts. However, neither scales well to large heterogeneous data sources nor adapts to different agent behaviors. In this paper, we introduce EvoOntology, a self-evolving ontology layer for data agents. EvoOntology encapsulates the ontology as an MCP server comprising a schema layer, a content layer, and a tool layer, enabling agents to actively query and interact with the ontology at runtime. To this end, we introduce a builder agent for autonomous ontology construction and a self-evolution loop that continuously refines the ontology through attribution-guided typed edits that are accepted only after a backbone-conditional paired evaluation. Experiments on three well-adopted data-agent benchmarks with four LLM backbones demonstrate that EvoOntology consistently outperforms strong baselines and existing semantic-layer approaches, effectively bridging the agent-data gap and enabling more effective interaction with heterogeneous data. Code: https://github.com/ruc-datalab/EvoOntology

Digital agents must often adapt to new environments whose interfaces, tools, and failure modes are not fully captured by pretrained models. We introduce \textbf{RSIAgent}, a training-free multi-agent framework for recursive self-improvement through autonomous memory construction. RSIAgent coordinates curriculum, actor, and verifier agents to continually explore the environment, validate outcomes, and retain environment-specific knowledge, including reusable causal relationships between actions, conditions, and consequences. It further adopts a \textbf{broad-then-deep} exploration strategy, combining parallel broad recursive self-exploration for discovering diverse environment structures with focused deep self-exploration for uncovering hard cases, hidden constraints, boundary conditions, and previously unknown causal dependencies. The resulting memory is frozen and can be directly reused for downstream tasks without updating model parameters. Experiments on OSWorld-v2 and Agent's Last Exam show that RSIAgent substantially improves strong open-source models, enabling Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6.

The Core is a unique competitive co-evolution algorithm that allows agents to evolve autonomous control without utilizing a traditional fitness function. The agents evolve via local interactions through tournament selection, crossover, and mutation, producing offspring by evolving better controllers. Previous works have shown The Core's ability to evolve agents capable of combat and navigation in the Xpilot video game. This research expands upon that premise by niching agents to specific subsets of the original environment The Core was tested in. Our results demonstrate the niched agents capacity for success over agents niched to the entire system and agents niched to different sub-environments.

Recursive self-improvement is becoming increasingly vital for autonomous AI agents, where progress hinges on discovering high-value solutions across complex domains. The driver of this process is effective exploration, however, managing and improving exploration strategies remains a major bottleneck. Current systems face a fundamental dilemma: fixed strategies fail to adapt as search spaces scale, while online policy optimization requires navigating vast meta-search spaces under delayed and expensive feedback over long-horizon rollouts. We introduce \textsc{Dream-RSI}, a framework for scalable and recursively self-improving exploration. A lightweight orchestration layer makes exploration explicit and programmable while leaving the underlying coding agent unchanged. Our key insight is that accumulated discovery history can serve as a replay simulator over the realized search space. By performing dreaming in the replay simulator constructed from historical discovery trees, \textsc{Dream-RSI} secures immediate, low-cost off-policy feedback to evaluate and refine exploration policies without invoking repetitive, expensive online evaluations. The improved policy is subsequently redeployed online to drive further discovery, continuously expanding the simulator pool in a self-improving loop. Across algorithm engineering, mathematical optimization, and GPU kernel engineering, \textsc{Dream-RSI} achieves competitive or improved discovery quality while substantially reducing discovery cost in several settings.

Domain practitioners understand their business constraints but may lack operations-research expertise or dedicated support. LLM-based optimization agents translate natural-language requirements into models or solver programs that established optimization tools can execute. This progress makes optimization more accessible, but real-world operations are dynamic: changing demand, resources, and priorities require updates to data, constraints, and objectives. Methods centered on isolated requests offer limited support for rapid adaptation that preserves earlier decisions and reuses useful search results. We introduce MAPLE (Memory-Augmented Planning with Language and Evolution), an agent for maintaining optimization problems through successive natural-language requests. MAPLE combines language-based problem construction with mathematical programming and evolutionary search. It retains the optimization program, accepted plans, earlier updates, and candidate solutions for subsequent requests. We introduce NLDO, a benchmark of 15 trajectories and 180 updates spanning selection, scheduling, rostering, routing, and cloud-resource placement. In the main evaluation, MAPLE completes all trajectories and achieves online scalar quality of 0.951 and a Pareto hypervolume ratio of 0.875. Controlled comparisons further show that maintaining executable state improves update validity and can preserve useful search information across substantial revisions.

Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing skill optimization methods often rely on costly execution-based evaluation and substantial task data. We introduce \textbf{COBRA-Skills}, an efficient framework that formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space. COBRA-Skills couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising or informative candidates while continually refining the skill population from execution feedback. Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance among compared methods, while reducing optimization cost by 55--58% relative to SkillOpt and using only 50 unique optimization examples per benchmark. Further analyses show that COBRA-Skills remains robust to changes in the agent harness and performs effectively when the target model itself is used for skill generation and refinement.

EGGROLL makes evolution strategies (ES) practical for LLMs by replacing dense Gaussian weight perturbations with low-rank Gaussian products, often of rank one. This choice is computationally attractive but geometrically severe: each rank-one perturbation lies in a zero-volume subset of the ambient matrix space, despite having identity covariance. We characterize the mean EGGROLL update field at finite rank and nonzero perturbation radii, then analyze the error of its finite-population estimator. The population field is obtained by applying an explicit resolvent to the gradient of the objective smoothed by the perturbations. We show that the resolvent can introduce a nonconservative component and can reverse the local stability of an optimum. EGGROLL is nevertheless exact on every quadratic objective at every rank and radius. For smooth objectives, its first local finite-rank correction is $O(\sigma^2/r)$, and nonasymptotic bounds control the resulting field error under smoothness assumptions. Under a local affine model, rank-one perturbations increase the variance of the gradient estimator by only $\frac{2(m+n+1)}{mn+1}$ relative to dense Gaussian ES, or $0.098\%$ for a $4096\times4096$ matrix. We then introduce LOO-ROLL, a leave-one-out estimator that preserves the finite-rank population field while replacing EGGROLL's two antithetic evaluations per direction by one. At equal evaluation cost, LOO-ROLL halves estimator MSE in transformer blocks. At matched wall time across ten post-training settings and models up to 8B parameters, LOO-ROLL improves seven outcomes in individual paired tests, with no significant loss. On the GSM8K test set, accuracy increases from $38.1\%$ to $63.0\%$ at 0.6B and from $65.9\%$ to $80.0\%$ at 8B. Transformer measurements recover the predicted finite-rank variance, while the rank comparisons show no reproducible reward-based advantage for rank eight.

How does cooperation evolve in complex agentic systems? Prior work in evolutionary game theory studies why individuals are incentivized to cooperate by isolating social interactions from the physical costs of behavior, while artificial life models traditionally study emergent self-replication without formalizing the dilemma between acquiring resources and preserving the shared energy needed to reproduce. In contrast, we introduce Autopoietic Game Theory, a computational model where social interactions, replication mechanisms, and their associated computational costs are endogenous and simultaneously co-evolving. We study these dynamics using a computational substrate of randomly initialized programs in Z80 machine code, showing empirically, and motivating with a simplified theoretical model, that embedding a social dilemma directly into the physics of computation can favor the emergence of self-replicating, cooperative strategies. When resources are scarce, our analysis shows that defection can become self-limiting even in well-mixed populations: parasitic stealing destroys shared energy, slows execution, and can prevent reliable replication. Empirically, evolved programs suppress stealing across several Z80 environments, while spatial assortment further supports structural complexity and task performance. We further show that the framework can incorporate exogenous pressures, such as math tasks structured as sequential social dilemmas, when rewards are tied to computation budgets. These results suggest that coupling an agent's capacity for computation to its available energy transforms cooperation into a dominant scaffolding for building sustainable, self-organizing systems.

Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.

Semantic-gradient-based prompt optimization (SGPO) improves agent harnesses using execution feedback, but its local update rule leaves the choice of edit scope and operation unresolved. We introduce RobustSGPO, which specifies the requested edit, constructs and checks the patch, and continues search from either the incumbent or retained snapshots. We evaluate permission scheduling, cumulative controls, and task-family transfer in the AgentX brainstorming workflow using 120 tasks, 95 runs, and 7,350 candidate attempts. Periodic $1\to2\to3$ scheduling exceeds fixed maximum permission by 0.28 test-score points. RobustSGPO increases completion on 30 held-out tasks from 60.0% to 80.0% and improves test quality from 3.77 to 4.14 under a 20-million-token budget. Category retention reduces source-task degradation after a shift, whereas random retention reaches a higher destination endpoint. Search-space control benefits quality through executable edits and alternative starting points, with measurable retention overhead.

Traditional software delivery assumes a static paradigm: code is constructed prior to execution and deployed as a fixed artifact. We present Agentic Just-In-Time Software Construction (A-JIT), a paradigm that replaces static binaries with dynamic, software systems that can perpetually evolve to meet changing demands. In A-JIT, an application is an integrated assembly comprising code, a runtime harness, and an embedded AI agent that continuously observes system usage and live execution traces. Much like a traditional JIT compiler specializes machine code to runtime execution paths, A-JIT specializes software logic, workflows, and tool interfaces to meet the specific needs of the end-user. By integrating synthesis directly into the ambient application lifecycle, A-JIT enables applications to dynamically construct missing implementations, generate new capabilities on the fly, and continuously adapt to end-user behavior. We demonstrate how this model supports trace-driven human-AI co-construction and opens a new design space for adaptive, self-evolving software.

Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. It can also repair a flawed expert prior. Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.

Agent skills provide a lightweight way to equip frozen language-model agents with domain knowledge and procedural guidance, yet obtaining high-quality skills remains costly and difficult to scale. Expert-written skills require substantial human effort. Recent skill self-evolution methods automate an iterative loop that uses execution feedback to revise skills, but their heuristic update strategies often yield unstable optimization and low iteration efficiency. We identify two challenges in realizing stable and efficient skill self-evolution. Direction Stability requires effective corrections to accumulate rather than be overwritten by iteration-local feedback. Update Adaptivity requires the scope of each revision to reflect the consistency of recent case-level improvements. We introduce SkillAdam, an Adam-inspired framework for optimizing discrete and non-differentiable skill documents. As a functional analogue of Adam's first moment, an optimization memory records identified problems and the outcomes of prior solution attempts to stabilize the update direction. As a functional analogue of Adam's second moment, a volatility-driven edit budget tracks the history-weighted variation of recent case-level improvements and adaptively controls the update magnitude. Across seven benchmarks that span short- and long-horizon tasks, SkillAdam achieves state-of-the-art performance with more stable optimization dynamics. It also obtains stronger skills with substantially fewer optimization iterations and lower cost than prior methods. Code repository: https://github.com/ruc-datalab/SkillAdam

Autonomous agents powered by large language models (LLMs) continuously accumulate experience through interaction, creating an opportunity to improve future behavior through self-evolution. A fundamental challenge is how to transform abundant, task-specific interaction experience into reusable model competence without sacrificing the ability to adapt rapidly to newly observed evidence. Explicit textual states, such as skills and agent harnesses, provide fast, human-readable and editable adaptation, but incur persistent dependence on external context; parametric policies provide compact and reusable competence, but are substantially slower to update. We present \textit{Experience Funnel}, a self-evolving framework that couples fast state adaptation with slow policy consolidation in an alternating loop. Interaction trajectories are first distilled into an explicit textual state, where newly acquired experience can be rapidly incorporated and validated. The framework then selectively identifies state-enabled behavior that remains useful across state revisions and consolidates it into the policy through transition-aware distillation. The updated state--policy pair subsequently generates new rollouts, providing fresh evidence for the next round of state adaptation and policy consolidation. Experiments across diverse agent benchmarks show that \textit{Experience Funnel} consistently improves agent capability over state-only evolution and policy-internalization approaches, while progressively converting useful explicit experience into autonomous policy competence.

The rapid growth of multimodal data has intensified the need for question answering (QA) systems capable of reasoning across heterogeneous sources such as text, tables, and images. In this paper, we present a comprehensive methodological comparison of three influential frameworks, namely Multimodal Adaptive Extraction (MAE), Solar, and UniMMQA, tracing the evolution of multimodal question answering from modality-adaptive pipelines to fully unified architectures. We examine how each approach models cross-modal interactions, transforms heterogeneous inputs, and performs reasoning, highlighting key design differences in modality representation, reasoning, and answer generation. Our analysis demonstrates a clear shift from explicit modality-specific processing toward unified text-centric formulations enabled by pre-trained language models (PLMs). Empirical comparisons across benchmark datasets show that this transition leads to substantial improvements in both Exact Match (EM) and F1-Scores, with UniMMQA achieving the most consistent and scalable performance. Despite these advances, we identify persistent challenges, including information loss during modality transformation, error propagation in multi-stage pipelines, and limitations in capturing fine-grained cross-modal dependencies. Overall, this study provides a deeper understanding of current design trends and offers insights into the future direction of unified multimodal reasoning systems.

Large language model (LLM)-powered agents can be accurate on average yet unreliable in production, a discrepancy that has been observed but remains largely unaddressed. When given the same task five times, a ReAct agent on the AppWorld benchmark using GPT-4.1 succeeds in all five runs only 53% of the time, even though its per-run pass rate averages 77%. We call this 24-point shortfall the consistency gap, and we argue that addressing it is a precondition for trustworthy AI agent deployment. We present a self-evolving agent framework that reduces this gap by identifying unstable, low-consistency steps in agent trajectories and converting them into episodic memory the agent can draw on in future runs. At its core is a Consistency Analyzer that pinpoints where and why a trajectory is likely to flip across executions, and a Guideline Generator that converts the diagnosis into targeted guidelines, committed to memory and injected into future agent executions on similar tasks. On AppWorld with ReAct/GPT-4.1, our framework raises the fraction of tasks that succeed in all five runs by +16 points on same-task evaluation and +13 points on similar-task generalization.

Multi-agent large language models solve complex tasks by coordinating several policies in a shared environment. However, existing reinforcement learning methods usually optimize each response or trajectory separately, even when several outputs jointly cause one state transition. Consequently, the update unit differs from the action executed by the system. To address this problem, we propose SRPO (Setwise Relative Policy Optimization), which treats the active set the minimal set of outputs consumed by one transition, as one multi-agent action. Specifically, SRPO combines member log-ratios into one cardinality-normalized set ratio, assigns one relative advantage, and clips the set once. This formulation unifies division of labor and joint co-evolution as actions with different set sizes. Experiments on mathematical reasoning and multi-turn search demonstrate one training interface for fixed, mixed, and dynamically routed workflows across four model scales, with the strongest macro-average results among the reported comparisons. Optimization diagnostics further characterize its stability under different event reductions and set sizes.

In persistent interactions, long contexts may encode an evolving process rather than a fixed record: later events can revise or revoke earlier information, changing what remains valid and what conclusions follow. We call this setting information evolution (IE). Solving IE requires identifying valid records, applying updates in order, and reconstructing the query-relevant state from the event history. Existing text-first synthesis pipelines make such data difficult to verify because state transitions and answer logic remain implicit. We introduce EvolveScaler, a code-driven framework that defines information evolution before rendering it as natural language. Human-authored operational specifications define state transitions, record validity, difficulty controls, and executable answer logic; a strong LLM then synthesizes a self-contained simulator from each specification. Executing validated simulators produces natural-language multi-turn event histories, while deterministic replay computes reference answers and atomic checklists. We instantiate EvolveScaler with 117 task prototypes and 159 final-question operators across five difficulty levels spanning approximately 7 to 1,200 events per instance, yielding about 35,100 training examples and 585 validated evaluation instances. On the very_long tier, the strongest model reaches 59.3% avg@5, while six models score below 10%. Training an internal A3B model on 6,000 EvolveScaler examples improves performance over its base checkpoint on all eight independently constructed out-of-distribution benchmarks, with a 5.25-point average gain. These results show that code-driven IE synthesis provides both challenging evaluation and transferable training supervision.

Large Language Models demonstrate remarkable proficiency in static reasoning, yet training them as autonomous agents through Reinforcement Learning (RL) for long-horizon tasks is often hindered by severe reward sparsity. While conventional \textit{agent-side warming} up via supervised fine-tuning (SFT) can alleviate this, it is frequently limited by data scarcity and constrained exploration. To address this, we propose a paradigm shift to \textit{environment-side adaptation} by constructing \textbf{F}eedback-\textbf{E}nriched \textbf{E}nvironments (\textbf{FEEs}). Through a pilot study, we establish a feedback design strategy that reformulates environments by transitioning from action guidance to observation enrichment during the later stages of both intra-episode exploration and inter-episode evolution. Large-scale experiments on SciWorld and BFCL benchmarks using various Qwen3 model scales and RL algorithms such as GRPO, GSPO, and DAPO demonstrate that FEEs consistently yield performance improvements over standard settings. Furthermore, our analysis reveals that training with FEEs \textbf{(1)} stabilizes training dynamics by reducing entropy volatility, \textbf{(2)} facilitates proactive state-space exploration in difficult tasks, \textbf{(3) }ensures the internalization of environmental guidance into policy weights rather than acting as a mere inference-time prior, and \textbf{(4) }identifies intra-group feedback consistency as a critical boundary for stable optimization.

Automating filament tracing in Cryo-Electron Microscopy (Cryo-EM) is essential for 3D helical reconstruction but challenged by intersecting topologies and extremely low Signal-to-Noise Ratios ($\text{SNR} = σ_s^2/σ_n^2$ < 0.1 or -10 dB). Existing paradigms fail: pixel-wise segmenters suffer from severe topological fracturing, box-based detectors face ghost center drift, sequential trackers derail due to error accumulation, and traditional active contours collapse under artificial closed-curve constraints. To resolve these bottlenecks, we present FPicker, the first topology-guided framework reconciling these incompatibilities. It unifies perception via a center-endpoint representation and an open-curve evolution module to explicitly model non-cyclic connectivity. On simulated benchmarks, FPicker outperforms top baselines by over $40\%$ relative gain in mean spatio-angular precision (mSAP) and reduces topological gap rates by over $60\%$ under extreme noise ($-20\text{ dB}$). By learning intrinsic physical geometry rather than local texture, FPicker demonstrates strong potential as a resilient geometric backbone. Its zero-shot performance on the real-world EMPIAR dataset exhibits robust topological resistance, achieving a state-of-the-art 82.9% mSAP upon fine-tuning. Our results also suggest modeling physical priors is a highly robust path toward bridging the sim-to-real gap in signal-starved scientific imaging. The code is publicly available at: https://github.com/tomzhaosky/FPicker.

Lifelong navigation (LN) requires an embodied agent to solve a sequence of navigation subtasks in the same environment. Since solving each subtask from scratch incurs redundant exploration, an LN agent must consolidate experience from earlier stages and reuse it in later stages, often through persistent scene representations such as scene graphs or visual snapshots. However, existing approaches typically assume a stationary environment, whereas in real-world LN settings, human activities can cause the environment to evolve. With the stationary assumption violated, existing methods may fuse outdated prior observations with new observations, yet current benchmarks cannot reveal this failure mode. In this paper, we present EvoNav-Bench, which extends the GOAT-Bench style LN formulation in the context of evolving environments. Built on the ProcTHOR framework, EvoNav-Bench introduces environment modifications between navigation tasks, making prior experience useful but not fully reliable. This design enables controlled evaluation of how environment evolution affects LN agents that reuse prior scene observations. Using EvoNav-Bench, we benchmark three recent methods that build and reuse scene representations for navigation. We also compare three simple heuristic strategies for handling environment evolution: Frontier-Update, Fail-then-Update, and Stage-Reset. Our results show that existing methods are brittle under environment evolution, while the heuristic strategies enable a controlled analysis of how agents can adapt to scene changes and mitigate their impact.

Temporal graph learning models the evolution of dynamic systems, where both structural interactions and semantic states change over time. However, existing benchmarks primarily emphasize structural evolution via temporal link prediction (TLP), while support for semantic evolution remains limited. Although temporal node classification (TNC) is sometimes included, it is typically restricted to simplistic binary settings that fail to capture realistic semantic drift. Moreover, commonly used datasets exhibit high link repetition, leading to inflated performance estimates and obscuring true model capability. To address these limitations, we introduce \textbf{TTGBench}, a new benchmark that jointly evaluates structural and semantic evolution. TTGBench comprises six real-world, text-rich datasets characterized by \emph{Dual Volatility}, enabling rigorous and fair evaluation of existing models. Notably, it is the first benchmark to support both multi-class and multi-label TNC, filling a critical gap in evaluating temporal semantic drift. We conduct a comprehensive evaluation of 17 state-of-the-art methods across Temporal Graph Neural Networks (TGNNs) and Large Language Model (LLM)-based paradigms. The results reveal a clear \emph{capability divide} between the two paradigms: TGNN-based methods excel at structural prediction but fail at semantic tracking, whereas LLM-based predictors show the opposite trend. Through in-depth analysis, we uncover their fundamental limitations and provide insights for developing more comprehensive temporal graph models.

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

Harness self-evolution is the process by which an agent modifies its prompts, tools, code, or orchestration in response to task feedback while keeping the underlying language model frozen, with changes persisting across subsequent tasks. We provide a systematic theoretical analysis of the feasibility and limits of safe harness self-evolution, connecting modification generation, finite-data certification and selection, safe adoption, and behavior after an update. Under a fixed user-task distribution, we establish conditions guaranteeing overall expected-reward improvement while controlling changes on retained tasks, characterize the probability of generating qualified modifications, and derive finite-data bounds for safe selection and adoption. Our analysis shows that generation and certification impose distinct constraints: current task performance does not determine the probability of generating qualified modifications, and generating more candidates need not improve the guarantee of a successful update when evaluation is limiting. Stagnation may therefore arise even when improvement opportunities remain. We further show that worst-case evaluation cost for recognizing genuine improvements diverges as expected reward approaches its upper bound. Across successive updates, certified improvement guarantees accumulate over a finite run, but a successful update does not by itself guarantee that further improvement remains possible. These results provide a basis for diagnosing bottlenecks and designing safer self-evolution mechanisms.

Language-model agents are vulnerable to indirect prompt injection (IPI) during tool use: adversarial instructions hidden in untrusted tool outputs can covertly redirect legitimate task execution. Existing work often trains and evaluates defenses against fixed attacks that do not adapt to the defender's behavior, so the resulting defenses may struggle against adaptive attacks. We combine adaptive attacker-defender co-training with subsequent refinement: continued interaction improves both roles, while learned attackers provide training challenges for further gains in defender safety and task utility. We therefore model adaptive IPI as a general-sum Markov game: the defender advances the task through successive tool calls, while the attacker can inject multiple times within the same task and adapt subsequent attacks to the defender's responses. Building on this formulation, we propose CoER, a verifier-grounded co-evolution and refinement framework. After initializing the attacker from successful trajectories, Co-PPO retains historical policies from both roles as opponent populations and mixes current and historical opponents for bilateral reinforcement learning, extending training beyond the latest matchup. Attackers from these populations are then reused to challenge teacher agents, and only demonstrations verified for both safety and task completion are used to fine-tune the co-evolved defender. In our main seven-domain evaluation, CoER reduces observed overall attack success from 38.5% to 0.2% and raises task utility from 63.2% to 76.3%, with improved attack resistance on external benchmarks.

Recursive self-improvement (RSI) lets a system improve the model-building machinery from its own failures, so every later model inherits the gain. Yet RSI has been validated almost exclusively on coding and formal benchmarks such as science QA and mathematics. This format bound limits RSI to improvement within a machine-checkable slice, not general capability where questions are open and correctness is settled by argument, replication, or measurement. We argue RSI must next operate across real, diverse scientific, engineering, and meta-scientific domains, not where formal evaluation is merely tractable. To that end we present MetaRSI-v1, where improvement is the scheduled composition of three typed operators over one unified paradigm. Data-RSI amplifies existing competence and marks its boundary; Harness-RSI edits a five-slot scaffold without touching weights; Model-RSI internalizes capability into parameters through bounded training. Sharing one loop kernel and artifact vocabulary, they make data, scaffold, and model changes composable rather than exclusive. A two-axis optimizer jointly decides operator order and each operator's proposal policy, while a meta-level policy revises the schedule across terms. We validate MetaRSI-v1 under the field's standard evaluations, on code and closed-form science, with no external teacher: the target model plays every role in its own loop. MetaRSI-v1 reframes self-improvement from a single-surface edit to a composition across the full model-production pipeline, opening two paths: a model route internalizing capability through training, and a harness route leaving weights untouched and thus extending self-improvement to any model reachable through an interface, with Data-RSI redefined as the shared substrate feeding both. The framework further yields refutable laws on where loops exist, how operators compose, and what supervision buys.

Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.

We present Dalek, a closed machine designed for agents that realizes self-maintenance, self-evolution, self-reproduction, and self-organization on any substrate satisfying a general host contract. The machine is built from three primitives---actors, messages, and channels. Four obligations---a host boundary, a construction language, admissible transitions, and rule heredity---give its boundary, identity, and closure a structural basis. Von Neumann's 1948 self-reproducing automaton supplies a hereditary constructional core: a self-description together with a constructor, a copier, and a controller. Dalek combines this core with the four obligations and rederives its medium for a text-and-message agent substrate, adding explicit structures for boundary, identity, history, and growth. A large language model and a compiler occupy the payload position and form a general capability producer. New capabilities are authored, compiled, installed into the description, and inherited by descendants. The same path produces the machine's own organs and even its runtime, closing heredity and evolution within the machine.

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.

The last five-plus years have seen many protein engineering disciplines transformed by advances in machine learning (ML), but the same cannot be said for directed evolution. Reflecting on a previously co-authored perspective, I discuss why I believe this to be the case, arguing that a disconnect between the goals of machine-learning-assisted directed evolution (MLDE) researchers--"identify an optimal protein"--and the goals of directed evolution more broadly--"identify a sufficient protein given time and resource constraints"--is a principal culprit. As an example, I highlight how nearly all current MLDE methods neglect to account for the cost of DNA synthesis, resulting in strategies that have limited practical applicability regardless of the underlying models' capabilities. I close by discussing recent works that are exceptions to this overarching trend, and emphasize that the last five years of efforts in ML-assisted protein engineering and the prescribed reframe of MLDE objectives need not be mutually exclusive.

Emerging edge AI workloads increasingly require arithmetic units that can trade computational accuracy for efficiency on demand. However, existing approximate arithmetic circuits are typically fixed-accuracy or rely on predefined structures for runtime configurability. This work introduces CircuitsDNA, an evolutionary framework that automatically evolves accuracy-configurable arithmetic circuits supporting multiple accuracy modes within a single circuit. It integrates three key features: 1) multi-threshold verifiability miter to enforce mode-specific accuracy requirements, 2) resource-limited verifiability-driven search to reduce verification overhead without sacrificing correctness, enabling efficient exploration of large circuit design, and 3) feedback-driven adaptive mutation to prioritize effective structural modifications and accelerate search convergence. Experimental results show that the 8-bit multiplier variants synthesized in 28-nm CMOS reduce the area-power product by up to 56% on INT8 DNN workload and 93% under exhaustive activity, compared with an exact 8-bit multiplier. Across CNNs and DeiTs, the accuracy loss relative to FP32 remains below 2% after fine-tuning under worst-case error (WCE) budgets of at most 1%. CircuitsDNA eliminates all search stalls observed in conventional methods across 8/12/16-bit multipliers, while adaptive mutation provides up to 1.33 times faster convergence than its non-adaptive counterpart.

Scientific agents must learn not only how to reason, but also what to believe. However, existing LLM agents typically express scientific hypotheses in free-form text, leaving their beliefs implicit and difficult to test or revise. We introduce EvoSCM, which equips scientific agents with explicit structural causal models that evolve as new experimental evidence is collected. EvoSCM maintains a population of competing SCM hypotheses, each encoding a candidate causal explanation of the environment, and evolves them through a closed discovery loop. In each round, the agent abduces latent mechanisms from accumulated evidence, designs discriminative interventions, and commits to falsifiable predictions that it tests through experimentation. Discrepancies between prediction and observation are inductively distilled into correction rules that revise the causal structures and mechanisms of each hypothesis, and the agent then deductively validates the revised population against accumulated evidence and structural consistency to guide the next round. We evaluate EvoSCM on DiscoverPhysics, a benchmark requiring agents to uncover the hidden dynamics of noncanonical physical worlds through experimentation. EvoSCM consistently improves scientific discovery over baselines, yielding more accurate explanations and predictions while making more effective use of experimental interactions.

The growing rate at which LLM agents interact with one another raises key questions about language evolution in multi-LLM-agent settings, with implications for safety and monitorability as well as for linguistic accounts of LLMs. To address these questions, we introduce GlossoGen, a novel platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, we build the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. We find that language evolution does occur between LLM agents, that the resulting languages are compositional and morphologically productive, and that they deviate from the LLMs' English prior in ways that render them incomprehensible to humans. Moreover, we identify several qualities essential to this evolution: pressure towards efficiency; the strength of the models backing the agents; and access to a "postmortem" stage in which agents can agree on linguistic conventions. Importantly, we observe that different conditions govern the transmission of language to new agents. Specifically, we find that agents learn new languages from usage alone, take an active role in this learning, and that while stronger models are required for novel language emergence, weaker models can learn an existing language once it has emerged. Taken together, our results indicate that current LLMs have the potential for cumulative cultural evolution -- previously attested only in humans -- with mixed populations of agents developing capacities that go beyond their lowest common denominator.

Protein inverse folding aims to recover amino acid sequences for a given 3D protein structure, underpinning broad applications such as enzyme engineering and drug discovery.Current methods often follow a serial pipeline, in which a structure encoder predicts a coarse sequence, which is then refined by protein language models (PLMs). However, because PLMs only perform post-hoc sequence edits, the refinement is bounded by the quality of upstream predictions.Thanks to recent multimodal protein language models (MPLMs), we could directly encode structure to generate sequences with pretrained structural knowledge, but we observe that they are not effective for inverse folding. Therefore, we introduce a symmetric dual-path architecture that both leverages PLMs for pretrained sequence evolution knowledge and MPLMs for pretrained structural knowledge to iteratively guide protein sequence generation.Through extensive experiments across standard protein inverse folding benchmarks, our method achieves state-of-the-art performance, surpassing prior approaches, and ablation studies validate the rationale of our symmetric design, revealing a promising direction for the community.

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.

Training open-ended agents via reinforcement learning (RL) is hindered by the lack of verifiable gold answers and scalable rubrics. Moreover, even near the model's capability boundary, long-horizon open-ended agentic tasks often yield brittle and unstable rewards, resulting in weak or noisy rollout contrast that obscures fine-grained optimization signals for group-based policy learning. To address these challenges, we propose ARISE-RL, a novel full-cycle self-evolution framework that couples a task/rubric Generator and a reasoning Solver through rubric-mediated co-evolution. The Generator grounds tool-related rubric criteria in real tool observations and is rewarded for producing valid, intermediate-difficulty tasks aligned with the Solver's evolving capability boundary. The Solver, in turn, learns from fine-grained rubric satisfaction signals through multi-step reasoning and tool use. We further introduce Reward-Gated Self-Evolution Distillation (RG-SED), which selectively distills a memory-augmented variant of the same policy back into itself only when the memory yields empirical reward improvement, thereby reducing distribution mismatch and avoiding blind imitation of noisy guidance. Finally, to support rigorous evaluation, we present ECR-Bench, an expert-calibrated rubric benchmark suite covering single-tool deep research and multi-tool travel planning. Extensive experiments demonstrate that ARISE-RL consistently achieves robust and stable overall state-of-the-art performance across all evaluated benchmarks.

Self-evolving agents advance toward autonomy by optimizing their harness---prompts, skills, tools, and execution logic---based on environmental feedback. This paradigm, however, is hampered by three challenges: \textit{credit assignment failure}, where terminal success/failure feedback makes it ambiguous which step caused the error; \textit{shortcut learning}, where agents memorize task-specific patterns rather than acquire generalizable capabilities; and \textit{catastrophic forgetting}, where unguarded updates degrade previously acquired competence. In this paper, we introduce HarnessEvolve, a self-evolving framework that learns from reference trajectories to achieve reliable agent self-evolution. HarnessEvolve decouples the execution agent from the evolutionary pipeline, assigning execution, evaluation, optimization, and gating to independent agent modules, enabling generalizable and stable harness improvements. Specifically, HarnessEvolve overcomes credit assignment failure by generating reference trajectories (execution paths produced when given the ground-truth answers) and aligning failed executions against them to extract error signals, which are clustered to reveal systematic failure patterns. To prevent shortcut learning and catastrophic forgetting, candidate harness updates must pass two gates: a quality gate that filters data leakage and prompt bloat, and a performance gate that accepts each update if it improves on the current batch without degrading recent batches, with epoch-end validation on a held-out set selecting the best-performing accepted agent snapshot. We conduct extensive experiments on several benchmarks spanning open-domain and enterprise scenarios, using different models and agent frameworks. Results demonstrate that HarnessEvolve consistently outperforms state-of-the-art baselines across all benchmarks and settings, confirming reliability across task domains.

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.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.

Language model agents increasingly propose actions, observe external feedback, and explain their own behavior. Their confidence and rationales are convenient monitoring signals, but convenience is not verification. We introduce an environment-grounded audit in which every intermediate proposal receives an exact outcome. A language model operates an evolutionary Contexto search whose feedback function assigns every valid guess an exact rank without human annotation. Across 200 runs spanning five configurations and three model families, four reporting configurations produce 12,249 self-reports. We test three assumptions: stated confidence is calibrated, inherited rationales affect later proposals, and fitness-based selection improves report quality. All three fail. Operators overstate top-100 success by factors of 4.8 to 9.3, while calibration and discrimination dissociate across model families. Controlled interventions on 754 inherited rationales bound any measured benefit of the genuine rationale to roughly 250 ranks. Neither fitness-based nor random selection produces a detectable selection differential or parent-to-offspring transmission in report accuracy, despite sharply different search behavior. Agent self-reports should therefore be treated as claims to verify against the environment, not as evidence of their own reliability.

Frontier language models that refuse harmful single-turn prompts often comply when the same intent is reached gradually over many turns, making multi-turn attacks one of the least understood failure modes of large language models. Most automated red-teaming methods treat this as a generation problem: produce attacks that break the model. We argue it is better framed as a search problem: discover, organize, and iteratively refine a diverse archive of attack strategies, producing a structured map of how a target model fails rather than a list of one-off successes. We introduce EvoFlint, which applies evolutionary quality-diversity search to multi-turn red-teaming. Attack strategies are phased conversation plans, not raw prompts, and are evolved through LLM-driven mutation and crossover. A Pareto fitness over attack success rate and peak severity preserves selection signal from near-miss attacks. A risk-indexed archive runs novelty search with local competition over strategy description embeddings inside each cell, maintaining diversity without committing to a predefined style taxonomy. A generation-level memory accumulates target-model insights across the population and feeds them back into strategy generation. On the HarmBench-test split, EvoFlint reaches attack success rates of 35.8% on Claude Sonnet 4.6, 59.7% on GPT-5.4, and 94.3% on Qwen3-32B, alongside 98.7% on the older GPT-4o included as a baseline reference. The resulting archive, organized by risk category, exposes for each target which categories of harm its safety training has and has not covered.

As people increasingly interact with LLM assistants in daily life, continually adapting to individual preferences has become essential for effective long-term interactions. However, user preferences are rarely stated in full, and instead emerge through heterogeneous, latent, and noisy signals, with existing methods relying on raw interaction histories or costly reward-based optimization to manage personalization. We introduce HypReflect, a reliable, scalable framework for continual personalization that infers explicit, uncertainty-aware preference hypotheses from diverse user signals, reflectively refines them as new evidence accumulates, and incorporates the resulting user model through hypotheses-guided self-distillation. Experiments across three personalization settings: online personalization, multi-session interactions, and implicit behavioral signals, show that HypReflect outperforms a range of baselines, including raw-history and incremental-update methods. We further demonstrate strong generalization to unseen users and cross-domain settings, along with stability across context budgets, reusable hypotheses, and more focused personalization. These results suggest a step towards reliable and scalable continual personalization through explicit, revisable user preference hypotheses.

AI is increasingly used in the R&D process that produces future AI systems. We study the conditions under which this feedback becomes self-amplifying. Our model describes how the rate of AI capability growth depends on baseline research productivity, recursive feedback, and the increasing difficulty of research progress. We derive a recursive reproduction number, $\mathcal{R}_{\mathrm{AI}}$, that determines whether improvements are amplified or damped across development cycles. This quantity compares the strength of feedback with the rate at which further progress becomes more difficult. When $\mathcal{R}_{\mathrm{AI}}>1$, the effects of improvements compound across development cycles, placing the system in a self-amplifying regime. When $\mathcal{R}_{\mathrm{AI}}<1$, their effects weaken across cycles. The transition depends on the structure of the AI R&D feedback loop and need not occur at any particular level of model capability. A system can therefore enter a self-amplifying regime before acceleration becomes visible, while rapid progress can also occur without self-amplification. Higher baseline research productivity can accelerate progress without changing whether the system is self-amplifying, but the duration of the development cycle becomes a limiting timescale for amplification. Increasing research difficulty can end a period of self-amplification. Extending the model to multiple research actors shows that improvements shared across organizations can make the overall research ecosystem self-amplifying even when no individual actor is. The framework identifies measurable properties of AI R&D systems that can help distinguish recursive amplification from rapid progress driven by other sources, including the strength of recursive feedback, how effectively improvements propagate into successor systems, cycle duration, and the increasing difficulty of further progress.

The performance of artificial intelligence (AI) and machine learning (ML) models degrades when the problem they were trained on drifts. This is a near-universal feature of real-world problems, which often change unpredictably. Biological evolution has achieved intelligence by overcoming this obstacle through natural selection acting on heritable variation. AI/ML techniques have long incorporated forms of natural selection, but it has been challenging to maintain model diversity as optimization naturally drives convergence. Here we show that a swarm of AI/ML models subjected to deliberate mutations of their model coefficients away from optimality can reliably and sustainably improve performance in changing environments by acting as a statistical hedge against non-stationarity. We call this mechanism 'Flawed in Nature, Perfect through Evolution', reflecting that the collective performance gain goes at the expense of individual performance. We prove via four theorems that the resulting regret reduction is guaranteed under general conditions, establishing the Flawed-in-Nature mechanism as a generalizable design principle for AI/ML systems. We validate these results on synthetic linear regression tasks, demonstrating that the mutated swarm delivers the best model in $\sim80\%$ of environment changes and that inference synthesis successfully translates this individual advantage into a collective one. The mechanism proves to be most effective when the mutation drift rate matches the drift rate of the environment. We outline a simple, adaptive controller that enables practical applications by tuning the mutation drift rate to match the unknown drift rate of the environment. The close analogy of the Flawed-in-Nature mechanism to biological evolution suggests it may have been a critical missing ingredient for the organic discovery of AI forms that more closely mimic biological intelligence.

Research planning is the decisive capability of AI scientists. Yet a research plan admits no verifiable answer, so reinforcement learning lacks the environment it requires: tasks paired with a critic. Rubrics extracted from scientific papers can supply the critic. Existing pipelines, however, draw the question and the criteria from the same content, so the reward can be earned by paraphrase. The rubric is further compressed into a single scalar per rollout. We introduce PaperGym, a unified framework that turns each research paper into a complete training environment. PaperGym exploits the structure of a paper: the question is synthesized from the research goal and background, while the criteria are derived from the method and experiments. The criteria span methodological innovation and experimental design, and criterion leakage falls to 3.7%, versus 11.90% to 34.10% in existing datasets. Training uses the rubric twice: first as privileged context for OPSD's self-teacher, then as the reward for GRPO. Across Qwen3-1.7B/4B/8B, this schedule outperforms supervised fine-tuning, either stage alone, and the reverse ordering, improving five-benchmark averages by +5.6, +5.0, and +4.8 points. With the recipe held fixed, models trained on PaperGym-20k win 58.1% of three-way comparisons, against 28.2% for RubricHub Science. The trained Qwen3-8B reaches 73.48 on ResearchQA, above the far larger Kimi K2.6. We release the pipeline, the 20,000-instance corpus PaperGym-20k, and the benchmarks PaperGym-Innov and PaperGym-Design.

Many important forms of human learning begin with a vague goal, such as "become a better physicist" or "improve at research." Learners must interpret the goal, identify capability gaps, decide how to learn, and determine whether they have actually improved. In contrast, existing work on LLM self-evolution typically begins with tasks and evaluation metrics specified by humans, reducing self-evolution to optimizing an explicit objective rather than deciding what and how to learn. We introduce ASPIRE, a benchmark for vague-goal-driven self-evolution. ASPIRE provides only a natural-language capability goal while downstream evaluation tasks remain hidden. The agent must operationalize the goal by choosing data and update methods, constructing training and validation signals, and deciding when to evaluate. ASPIRE supports both model-weight and agent-harness evolution in a unified interactive environment and evaluates the resulting systems on a hidden, expert-authored set of 520 items spanning six goals. Our experiments show that vague goals redirect search effort toward goal interpretation. Current agents routinely complete training and harness-editing loops, but weight-level gains remain sparse and unstable, and the strongest evolved harness remains below the engineered Qwen-Agent reference. Agents often train on mismatched data and trust narrow self-evaluations, so local gains fail to transfer to hidden evaluation and continued search and training can erase earlier improvements.

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.

On-policy distillation (OPD) offers dense token-level supervision as an alternative to the sparse outcome-level advantages of reinforcement learning with verifiable rewards (RLVR). However, the teacher scores student-generated trajectories that are inherently off-policy for it, so the reliability of its supervision, and hence the source of the student's improvement, remains unclear. We quantitatively analyze teacher supervision during OPD training and find substantial noise whose prevalence increases with teacher scale. Surprisingly, the student policy is insensitive to such noise, converging to comparable performance regardless of whether noisy supervision is retained or removed. Does OPD distill at all? By analyzing what drives its gains, we find that learning concentrates on low log-probability tokens, and using a single fixed negative advantage matches the performance of teacher-provided ones. This suggests that OPD works largely by suppressing low log-probability tokens, which requires no teacher. These findings motivate On-Policy Self-Adaptation (OPSA), a supervision-free method using entropy-adaptive negative advantages. It assigns stronger learning signals to high-entropy positions, suppressing tail tokens, and evenly redistributing probability mass among head tokens. Compared with the base \texttt{Qwen3-1.7B}, OPSA improves Avg@32 by 35.41 points on AIME24, corresponding to a 263% relative gain, and more than doubles Pass@32 across all three benchmarks. It also outperforms OPD by 16.77 points in Avg@32 on AIME24. Extensive experiments and analyses across model families and tasks further demonstrate its effectiveness and generalizability.

Long-horizon complex tasks require foundation models to accumulate information, maintain internal states, and adapt over extended interactions. Safety should be an intrinsic property of the model itself, rather than a behavioral constraint relying solely on external safeguards or post-hoc alignment such as supervised fine-tuning. This motivates Safety from Within, where safety-relevant capabilities are represented and invoked through the model's native computation. We present Safin-1, a family of foundation models realizing this principle through memory routing and state evolution. Safin-1 is built on Memory-Anchor Routing across Context History (MARCH), a network architecture that maintains structured memory states and selectively retrieves relevant historical information through content-conditioned routing. It supports test-time adaptation of persistent capability states without repeatedly modifying the backbone, enabling controlled specialization over a shared foundation. We investigate this interface on downstream safety tasks through a Safety State, demonstrating effective state-based adaptation with substantial safety improvements. More broadly, the routed-state interface unifies contextual memory and persistent capability adaptation within the model's native computation, reframing memory from a passive record of prior context into an active substrate for maintaining and evolving model behavior. Evaluations across general capabilities, long-context understanding, retrieval, and efficiency further validate Safin-1. These findings provide a path toward safety as a state-native and adaptively maintainable capability. This work is only an initial architectural exploration of Safety from Within, and substantial further work is needed to realize this broader vision.

Recent studies have shown that multimodal large language models (MLLMs) can serve as embodied agents, translating language instructions and visual observations into executable plans. However, building agents that can continually improve through interaction and rapidly adapt to their environments remains challenging. Summing up experience from past interaction trajectories provides a promising solution, but existing experience-based methods often rely on manually designed prompting workflows to extract and update skills. Such fixed procedures may struggle to learn updated skills from new and diverse experiences. We introduce PRACTICE, which trains a skill learner to discover and maintain a persistent skill library from past interaction trajectories while keeping the task executor frozen. Given the historical accumulated skills and incoming trajectories, the skill learner produces structured batch-edits that add, refine, merge, or remove skills, and then hierarchical consolidate all collected edits into a consistent updated skill library. We train the learner with a two-stage curriculum. First, it learns basic skill generation and library maintenance from oracle trajectories. Then, by contrasting successful and failed trajectories from heterogeneous executors on the same tasks, it learn to identify invalid action patterns and recovery strategies. Finally, we apply online skill-edit distillation to align the skill learner with a stronger teacher on its current edit distribution to further improves the policy. Experiments demonstrate that a compact skill learner delivers consistent performance improvements across successive library-update rounds for multiple frozen executors. On EB-ALFRED and EB-Habitat, PRACTICE further outperforms the strongest experience-based baselines. Project resources are publicly available at: https://baai-agents.github.io/PRACTICE

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.

LLM-based agent systems increasingly adopt skill-based architectures to reduce repetitive reasoning costs and improve stable, efficient task execution. Recent studies propose self-evolving agents that autonomously generate, refine, and reuse skills from past experiences to enable continuous capability evolution. However, autonomous skill evolution introduces a new attack surface in which malicious capabilities are generated, stored, and reused as legitimate skills. In this paper, we define EvoSkill Injection as a threat model targeting the autonomous skill generation and evolution pipeline of self-evolving agents. We further propose SARGE (Red-teaming Autonomous Skill Generation and Evolution in self-evolving agents), a red-teaming framework for evaluating this threat model through iterative generation, escalation, and reinforcement interactions. To support our framework, we construct EvoSkillBench, a benchmark dataset of malicious interaction trajectories for inducing malicious skill formation in self-evolving agents, and introduce EvoSkillSafetyBench, a post-attack benchmark for evaluating whether injected malicious skills are subsequently retrieved and activated as harmful behaviors. Our evaluation shows that SARGE induces malicious skill formation and that injected skills are persistently stored and repeatedly activated, highlighting the risk of persistent capability corruption.

Direct vision-language-action policies generate continuous robot actions efficiently, but standard behavior cloning leaves two complementary gaps: their representations are not explicitly required to describe how the scene evolves over multiple time scales, and deployment trajectories of unequal quality are often reused without separating useful dynamics from undesirable behavior. We introduce \method, a direct world-action policy that combines outcome-agnostic predictive learning with outcome-aware policy improvement. \method first retains a local fixed-offset JEPA objective and adds trajectory-relative multi-horizon transition alignment at 25%, 50%, 75%, and 100% of the remaining episode. These training-only targets require the current policy representation to preserve both local physical changes and longer-range task progress, without supplying explicit future tokens to the action head. \method then trains an independent distributional value critic on cumulative deployment trajectories, computes action-chunk-aligned $N$-step advantages, and converts them into positive, negative, or null text conditions for a flow-matching actor. Thus, every valid trajectory can teach what physically happened, while the actor is deployed only under the condition associated with relatively better actions. The multi-horizon predictor and critic are removed from online execution, preserving direct action generation from the current observation, language instruction, and proprioception. \redclaim{Across the three simulation benchmarks, \method achieves the strongest overall performance while preserving the direct actor's online execution path.}

Computer use agents (CUAs) are vision-language models that perceive a screen and act on a real operating system through mouse, keyboard, and terminal, and they are increasingly deployed to automate everyday digital tasks. Because they can be exposed to untrusted content while operating, they are vulnerable to indirect prompt injection (IPI), in which an adversary plants instructions in content the agent will read and redirects it toward actions that violate the user's intent. Existing CUA safety benchmarks evaluate fixed injections written by hand, which may underestimate the risk posed by an adaptive adversary. We present SIR, a black box IPI attack that (i) composes stealthy injections from a small library of reusable principles stated in plain language and (ii) wraps composition in an iterative feedback loop that diagnoses the victim's failed trajectories and distills the bypasses into new, named strategies that are reapplied across tasks. Unlike prior red teaming of web agents, we target CUAs at the operating system level and score attacks with a fully deterministic oracle, using checks on filesystem, service, and permission state rather than an LLM judge. On experiment, we evaluate three frontier CUAs. Composing principles with feedback raises the attack success rate over a baseline written by hand, for example from 4% to 24% on Claude Opus 4.8 and from 0% to 28% on Gemini 3.5 Flash, while the benign task still completes. Principles discovered against one model further transfer to a different architecture with no additional feedback.

Large language models are increasingly required to generate responses that satisfy multiple competing objectives. Since optimal trade-offs depend on both user preferences and input prompts, controllable multi-objective generation must dynamically adapt models at inference time without retraining. To address this, we propose Evolutionary Soups, a mixture-of-experts framework for fine-grained generation control, with gating networks trained via an evolutionary algorithm. The per-layer gating networks dynamically produce expert-merging coefficients from hidden-state representations, while the evolutionary algorithm incorporates greedy hypervolume contribution for effective evolution of these gating networks, achieving consistent improvements on large and noisy training datasets and broader coverage of the non-convex Pareto front. Experiments across three tasks demonstrate the effectiveness of Evolutionary Soups over baselines: it achieves the best hypervolume, linear utility, and Tchebyshev utility (~20% improvement) among controllable methods on all tasks.

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.

Communication typically relies on a shared code, and any change to it must be coordinated between senders and receivers to avoid a breakdown of communication. The honeybee waggle dance illustrates this problem: species with horizontal combs point directly at a food source, while species with vertical combs cannot point directly and instead reference the dance to gravity, decoded against the position of the sun. We model the rise of the first of these codes and its evolutionary transition to the second in populations of bee-like agents, with selection acting at the level of colonies. In a horizontal-comb model, we find that direct pointing evolves readily when food is moderately hard to find by random search alone, whether because sites are few and large or many and small, and fails when food is too sparse to spark dances or so abundant that it is found without signaling. Adding an exogenous benefit for vertical combs, we then find that the transition to the gravity-referenced code is driven mainly by the mutation rate and the magnitude of this benefit, with the coupling between sender and receiver mutations playing a further role at low mutation rates. Given a favorable confluence of these factors, the transition proceeds reliably and without a breakdown of communication.

Agent capability is not determined by the model alone. The agent harness, encompassing memory management, planning strategy, action protocol, and tool/skill orchestration, can dominate the contribution of the underlying foundation model. Yet harness design remains manual, task-specific, and fundamentally unscalable. We present JIT-Agent, a harness intelligence model trained to synthesize task-adaptive agent harnesses on the fly for arbitrary off-the-shelf agentic LLMs. We formalize the agent harness as a composable, machine-generatable artifact governed by a fixed four-module protocol, and train JIT-Agent to customize harnesses for a given task at hand, repair harnesses for stable and reliable execution, and self-evolve by distilling performance signals from an expanding archive of prior harness configurations. Equipped with JIT-Agent as a harness helper, DeepSeek-V4-Flash surpasses GPT-5.6 on DeepSearchQA (+9.1) and OdysseyBench (+4.3), while the already strong GLM-5.2 gains up to +20.2 points. Across controlled evaluations, JIT-Agent-generated harnesses are performance-competitive with mature agent runtimes such as OpenCode and Claude Code and consistently improve multi-scale model families of DeepSeek V4, Mimo-V2.5, and Qwen3.6. To our knowledge, JIT-Agent is the first model purpose-built for just-in-time harness generation, establishing harness intelligence as a trainable, transferable, and compounding dimension of agent capability orthogonal to model scaling.