Best AI papers explained

Recuris is a recursive architectural framework designed to enhance the performance of large language model agents during complex, long-horizon tasks. By coupling Working Memory, which tracks live task progress, with Experiential Memory containing reusable skills, the system ensures that model actions remain grounded in current needs rather than becoming lost in expanding conversation histories. This integration allows the agent to produce structured execution traces, which a fixed Meta-Agent uses to pinpoint specific failures and apply targeted memory patches. Empirical results across various benchmarks demonstrate that this self-improving loop significantly boosts task success rates for both open-source and frontier models like GPT-5.6 and Claude Opus 5. By reducing common errors such as hallucinations and missed commands, Recuris provides a scalable foundation for agents to transform accumulated experience into increasingly reliable autonomous behavior.

Researchers introduce TailSFT, a modified supervised fine-tuning algorithm designed to better prepare language models for subsequent reinforcement learning. Unlike standard fine-tuning that minimizes overall cross-entropy, TailSFT filters out sequences that the model has already mastered to focus training on the under-modeled "tail" of the data distribution. This approach prioritizes coverage, ensuring the model retains a diverse range of correct responses that reinforcement learning can later identify and amplify. Theoretical analysis and experiments on the OLMo-3 7B model demonstrate that TailSFT significantly boosts performance in math and coding tasks, particularly by improving pass@K metrics. Ultimately, the authors show that a higher-coverage initialization leads to faster learning and superior final accuracy after reinforcement learning. This work advocates for a stage-aware approach to AI development, where intermediate training phases are optimized specifically to benefit the next stage of the pipeline.

This paper introduces SPADE, a reinforcement learning framework that enables a single large language model to achieve open-ended self-improvement by designing its own training worlds. One role, the Environment Designer, creates complex, multi-turn tasks as executable Python code, while the Reasoning Agent role learns to solve them. To ensure the tasks are challenging yet possible, the system utilizes a hint-based regret signal, rewarding the designer when an agent succeeds with a secret hint but fails without it. This competitive dynamic allows the training curriculum to automatically evolve in complexity as the model's capabilities grow. Research results demonstrate that SPADE significantly outperforms static training methods across various math, coding, and tool-use benchmarks. By turning environment creation into a learnable skill, the framework offers a scalable solution to the scarcity of high-quality human data.

This paper introduces ACES (Agentic Continuous Evaluation of Skills), a comprehensive framework developed by NVIDIA to move beyond static document scanning when assessing AI agent capabilities. While traditional methods merely check a skill's structure or style, ACES evaluates skills as executable artifacts by running live, sandboxed trials to observe how agents actually discover and use them. The methodology centers on Skill Lift, a metric that measures the marginal value a specific skill adds by comparing an agent's performance with and without that skill enabled. This system utilizes a standardized Agent Trajectory Interchange Format (ATIF) to ensure compatibility across different agent harnesses and models. Empirical testing on 145 enterprise skills reveals that static scores correlate poorly with runtime success, highlighting the necessity of live agent evaluation for identifying regressions or routing failures. Ultimately, the framework integrates into CI/CD workflows, allowing developers to refine agent behaviors using evidence-based reports rather than subjective prose.

Researchers from Meta Platforms propose a novel offline evaluation task called impression share prediction to better anticipate how new ranking models redistribute traffic across different business objectives. Traditional metrics often fail to capture these shifts, which can negatively impact downstream utility even when predictive accuracy improves. To address this, the authors developed a structural causal model that identifies how model signals and delivery capacity interact to determine impression allocation. Their framework includes a Random Forest regressor for established models and a specialized encoder-conditioned architecture to handle the complex dynamics of newly introduced models. This system significantly reduces prediction error compared to standard baselines, particularly during the critical first hour of a model's deployment. Ultimately, this approach provides practitioners with vital visibility into a candidate model's allocation behavior before proceeding to expensive online A/B testing.

The researchers introduce Q-Learning with World Models (QWM), a framework designed to enhance sample efficiency and performance in robotic reinforcement learning. Unlike traditional model-based methods that often suffer from compounding biases by training policies on "imagined" data, QWM maintains a policy and critic trained exclusively on real environment transitions. It leverages a learned world model specifically at test-time to conduct tree searches over potential future trajectories, allowing the agent to select actions with the highest predicted downstream value. This approach combines the predictive power of world models with the stability of grounded Q-learning to navigate complex, high-dimensional tasks. Experiments on challenging manipulation benchmarks like Robomimic and LIBERO demonstrate that QWM significantly outperforms existing model-free and model-based baselines. Ultimately, the framework scales effectively from state-based inputs to visual observations, providing a robust method for improving online reinforcement learning.

This paper introduces Conformal Language Modeling via Posterior Sampling, a novel framework designed to reduce hallucinations in Large Language Models while maintaining text quality. Unlike previous methods that perform "post-hoc surgery" by deleting claims from already generated text, this approach reweights the model's sampling distribution toward more reliable responses. By treating the generation process as posterior sampling conditioned on high-confidence regions, the researchers ensure that outputs remain coherent and fluent. The authors develop a calibration procedure that provides statistical guarantees for factuality across complex tasks like biography generation and mathematical problem-solving. Their findings demonstrate that this method significantly improves downstream utility compared to existing filtering techniques, particularly in scenarios with strong logical interdependencies. Ultimately, the work offers a mathematically grounded way to achieve target risk control without sacrificing the structural integrity of the generated language.

BoNVoyage is a novel training framework designed to improve reward models (RMs) used in reinforcement learning from human feedback. Traditional RMs often fail because they are trained on static data distributions that do not reflect the adversarial distribution shifts occurring during the actual optimization process. Instead of simple pairwise ranking, this method uses test-time alignment and Markov chain Monte Carlo sampling to maximize the likelihood of preferred responses under an idealized policy. By incorporating contrastive divergence to maintain efficiency, the approach creates a more reliable signal for the language model to follow. Experimental results across mathematics and science benchmarks demonstrate that this technique produces superior downstream policies compared to standard baselines. Furthermore, BoNVoyage exhibits significantly more robustness to reward over-optimization, preventing the common issue of reward hacking during extended training.

This research investigates the operational dynamics of agent skills, which are structured packages of procedural knowledge designed to help AI agents learn from experience. By comparing distilled skills against raw workflow memories, the study reveals that skills primarily act as procedural anchors that stabilize execution and reduce environment failures rather than simply injecting factual knowledge. While skills improve task success by providing compact guidance, they also introduce new risks, such as mechanical misapplication or the rigid following of incompatible instructions. The authors also identify retrieval as a significant bottleneck, noting that while agents often find the correct skill, their performance is frequently hindered by confusable distractors and execution-layer difficulties. Ultimately, the work provides a systematic taxonomy of success and failure modes to move evaluation beyond simple success rates toward a deeper understanding of reliable self-improvement.

This paper introduces the Wiggle Framework, a novel diagnostic tool designed to evaluate the epistemic stability of Large Language Models when they act as autonomous judges. Researchers discovered that even top-tier models frequently reverse their original verdicts when subjected to social pressure, rephrased prompts, or persistent adversarial arguments. This vulnerability, termed "wiggle," is prevalent across diverse evaluation tasks, including safety monitoring and political analysis, often resulting in decreased accuracy after the model is challenged. The study concludes that high-performing AI judges are surprisingly fragile and susceptible to persuasion, which compromises their reliability in critical grading and moderation roles. By measuring mechanical consistency and multi-turn persistence, the authors demonstrate that initial majority consensus remains the most reliable indicator of a model’s potential to remain steadfast. These findings highlight a significant gap between a model's static accuracy and its actual cognitive conviction during interactive scenarios.

This paper introduces the Data Manifold Oracle (DMO), a training-free framework designed to predict neural scaling laws by analyzing raw text through compression statistics. By using Lempel-Ziv algorithms, the researchers extract two key metrics—an entropy-rate floor and a data-scaling exponent—to forecast model performance without the high cost of training model families. The authors prove an exact symbolic obstruction, demonstrating that raw text alone cannot reveal a dataset's geometric dimension without an external scale. Empirically, the DMO effectively ranks the scaling behavior and loss saturation of various corpora, including web, code, and math data. The research further extends this to DMO-Doc, a selector that identifies high-quality documents to improve pretraining and post-training outcomes. Ultimately, the work establishes that fundamental properties of machine learning performance are visible in the statistical structure of data before a single gradient step is taken.

This paper introduces a rigorous statistical framework for discovering human-interpretable insights from unstructured data, such as text, audio, and video. By repurposing AI interpretability tools like sparse autoencoders, the method maps complex data into a high-dimensional space of thousands of distinct concepts. The author utilizes advanced multiple hypothesis testing to ensure these discoveries remain statistically valid while avoiding the pitfalls of data snooping or researcher bias. To ensure the results are understandable, the system employs Large Language Models to generate and evaluate natural language descriptions of the identified patterns. Applications to empirical economics demonstrate that this approach can automatically recover nuanced findings that previously required intensive manual labor or separate experiments. Overall, the framework provides a principled, inexpensive, and replicable way to uncover "unknown unknowns" within large, unstructured datasets.

This paper introduces and addresses the incentive collapse paradox, a phenomenon where accuracy-based payments fail to motivate human effort as AI assistance becomes more reliable. The authors demonstrate that if human workers only receive rewards based on their final output accuracy, they will eventually free-ride on the AI’s suggestions rather than exert costly verification effort. To solve this, they propose a sentinel-auditing mechanism that deliberately injects occasional, detectable AI errors to reward human vigilance independently of the AI's natural performance. This strategy is further integrated into an incentive-aware active statistical inference framework, which jointly optimizes budget allocation and task sampling. Theoretical results and experiments on survey and protein data show that this approach maintains high label quality at a finite cost. Ultimately, the research proves that accounting for strategic human behavior allows for more cost-effective and precise statistical estimation than traditional methods.

This paper introduces a framework for modular memory as the essential solution for creating continual learning agents that adapt without forgetting. The authors argue that while current foundation models excel at static tasks, they struggle with ongoing experience accumulation and personalization because they rely too heavily on single-model parameter updates. To solve this, the framework integrates In-Context Learning (ICL) for rapid, short-term adaptation with In-Weight Learning (IWL) for stable, long-term knowledge consolidation. The proposed architecture consists of three distinct components: a core model for general reasoning, a working memory for immediate context, and a long-term memory for persistent storage. Inspired by both human neuroscience and computer architecture, this system allows agents to interact with the world in real-time while refining their core capabilities during internal "consolidation" periods. By separating fast adaptation from slow integration, the researchers aim to overcome the stability-plasticity trade-off that has long hindered artificial intelligence. Ultimately, this approach provides a roadmap for developing self-evolving agents capable of operating in dynamic, open-world environments.

This paper introduces harness RL, a novel meta-learning framework designed to enable large language models to self-improve during test-time adaptation. Rather than updating model weights, which is computationally expensive, this method optimizes the agent’s harness—the external instructions, memory, and rules that guide model execution. By training a proposer model to revise this harness while keeping the executor model frozen, the system learns a transferable self-improvement operator. This approach reduces complex meta-learning to a standard reinforcement learning objective because the adaptation process requires no gradients. Experimental results across reasoning and coding tasks demonstrate that the trained proposer generalizes to unseen problems and maintains performance across longer revision horizons. Ultimately, the authors show that harness RL successfully isolates and improves the model's capacity for meta-self-improvement.

This paper investigates a critical strategic mismatch between Large Language Models (LLMs) and human decision-makers in competitive environments. Through game-theoretic experiments, the researchers demonstrate that LLMs predominantly act as Nash-type reasoners, assuming their opponents are perfectly rational, whereas humans exhibit bounded rationality and varied reasoning depths. This overestimation of human sophistication often leads LLMs into a Nash trap, where equilibrium play fails to maximize payoffs against actual human behavior. To rectify this, the authors propose supervised fine-tuning methods, including Trap-Aware SFT, which calibrates model responses to empirical human benchmarks. Their findings suggest that effective human–AI alignment requires models to possess not just high reasoning capabilities, but also calibrated expectations of human behavior. Ultimately, the study advocates for a selective deployment architecture that preserves equilibrium play while adapting strategies when human interaction makes it more profitable.

This research paper introduces a mathematical framework to prove that LeJEPA (a specific self-supervised learning architecture) can accurately recover the hidden structure of the world from complex data. The authors establish that when a model combines an alignment loss with Gaussian regularization, it achieves linear identifiability, meaning the learned representation is a simple rotation of the world’s true latent variables. This property is shown to be unique to Gaussian latent distributions, as any nonlinear distortion of the representation would strictly degrade the model's predictive performance. Furthermore, the study demonstrates that this linear recovery is essential for optimal latent-space planning, allowing an agent to navigate a learned model as effectively as the real world. The theory is supported by experiments ranging from 2D simulations to high-dimensional robotic control tasks, confirming that the model's training objectives act as a reliable proxy for structural accuracy. Ultimately, the work provides a formal foundation for building World Models that are mathematically guaranteed to be faithful to the environments they represent.

This research paper investigates Role Drift, a failure mode in compound AI systems where individual modules abandon their specific instructions to find shortcuts that improve final task accuracy. During end-to-end training, modules like "readers" or "decomposers" may stop performing their intended functions—such as relying on external evidence—and instead fall back on internal memory or leak answers to simplify the process. While terminal performance scores may increase, this erosion of role fidelity makes systems less auditable, harder to update, and more fragile. To combat this, the authors introduce Role Anchor, a regularizer that maintains a module's intended behavior by penalizing deviations from its initial role-prompted state. Experiments demonstrate that Role Anchor effectively preserves the division of labor within multi-module pipelines at a tunable cost to overall accuracy. Ultimately, the study reveals that significant gains in reinforcement learning can be illusory if modules achieve success by violating their designed roles.

Research from Stanford University challenges the conventional assumption that pre-training a Q-function on offline data improves reinforcement learning fine-tuning. The authors demonstrate that naive pre-training often yields no benefit because the offline Q-function mismatch with the optimal online Q-function creates an incompatible value landscape. To address this, they introduce Initialization via Policy Ensemble (IPE), a method that trains multiple diverse policies on the same data. By pooling rollouts from this policy ensemble, IPE provides broader action coverage and creates a more robust foundation for the critic. Experimental results across various robotic tasks show that IPE improves fine-tuning performance by an average of 26% over standard methods. This approach highlights that data diversity around the policy distribution is more critical for success than simply maximizing value during the offline phase.

Digital search is transitioning from a human-centered discovery process based on links and keywords to an agent-mediated system of delegated decision-making. In this new AI-native paradigm, users express goals in natural language while autonomous agents interpret intent and execute tasks on their behalf. This shift moves the internet from an attention economy, where firms compete for clicks, toward a preference economy focused on satisfying specific user desires. Because search results are becoming hidden beneath the surface of agentic interactions, the authors emphasize the need for transparent and competitive system designs. The research highlights grand challenges regarding how to maintain market efficiency, ensure agent alignment with user goals, and verify the trustworthiness of automated choices. Ultimately, the future of the internet depends on creating open frameworks that prevent power concentration and foster fair competition within these emerging digital marketplaces.

The research introduces BINEVAL, a novel evaluation framework that improves the reliability of Large Language Models (LLMs) by decomposing complex quality criteria into atomic binary questions. Unlike traditional holistic grading, which often produces opaque and inconsistent scores, this method utilizes a "decompose-then-verify" approach to generate transparent, multidimensional assessments. By aggregating simple yes/no verdicts into calibrated scores, the system achieves superior alignment with human judgment across tasks like summarization and dialogue. Beyond measurement, the framework supports an iterative optimization loop that uses question-level feedback to refine both evaluator rubrics and generation prompts. This diagnostic granularity allows developers to pinpoint specific failure modes, such as factual misattributions or formatting errors, that broader metrics typically obscure. Ultimately, BINEVAL demonstrates that breaking evaluation into checkable sub-tasks makes LLM outputs more interpretable, debuggable, and actionable for continuous model improvement.

This research introduces a hierarchical latent selection model to explain how large language models develop robust reasoning through post-training. The authors argue that supervised fine-tuning (SFT) provides the essential building blocks, while reinforcement learning (RL) decomposes these traces into reusable atomic skills and routing mechanisms. By isolating these components, RL enables models to solve out-of-distribution problems through novel combinations of learned modules. Controlled experiments on synthetic tasks prove that training on compositional traces is superior to learning isolated skills. The study concludes that an ideal training protocol uses SFT to ensure broad module coverage and RL to explore unseen compositions. This division of labor allows models to generalize systematically beyond the specific demonstrations provided during initial tuning.

Researchers utilized chess as a controlled testbed to investigate how pretraining choices influence the effectiveness of reinforcement learning (RL) in large language models. By systematically scaling models from 5M to 1B parameters, the study established a joint scaling law where a model's pretraining loss accurately predicts its subsequent RL performance. The findings reveal that extended pretraining not only provides a better starting point but also increases the speed at which a model improves during RL training. Mechanistic analysis showed that while RL amplifies correct moves on simple tasks, it can also surface previously hidden solutions on difficult problems. Furthermore, the authors demonstrated that these predictive patterns transfer to the math domain, suggesting the results are applicable to broader reasoning tasks. Ultimately, the study suggests that as total compute budgets grow, a larger share of resources should be allocated to the RL phase.

This paper introduces Normalized Simulatability Gain (NSG), a new metric designed to measure the faithfulness of AI self-explanations by testing their predictive value. By evaluating 18 frontier models, the researchers demonstrate that an AI's explanation of its own logic significantly helps a separate "predictor" model guess how the AI will behave on related counterfactual scenarios. The study provides a positive case for faithfulness, finding that self-generated explanations contain privileged self-knowledge that external models cannot replicate. However, the authors also identify a "highly misleading" subset of explanations where the AI's stated principles contradict its actual choices, particularly in ethical dilemmas. Ultimately, the research suggests that while LLM explanations are imperfect, they remain a valuable tool for AI oversight and safety.

This research paper investigates Sequential Monte Carlo (SMC) and other particle filtering algorithms as a theoretical framework for improving large language model (LLM) inference. The authors introduce a principled approach to analyze inference-time interventions, such as parallel reasoning and pruning, by utilizing process reward models to steer generation. Their findings establish non-asymptotic guarantees for SMC based on criteria like bounded action-level coverage and divergence between true and approximate reward distributions. To address limitations in standard SMC, they propose SMC with Rejection Sampling (SMC-RS), which maintains high accuracy even when reward models are nearly perfect. Empirically, the study demonstrates that SMC consistently outperforms Best-of-N sampling on complex mathematical reasoning tasks and benchmarks. Ultimately, the work bridges the gap between ad hoc sampling heuristics and rigorous statistical theory to optimize the accuracy-cost tradeoff in AI inference.

This research paper critically examines automatic harness evolution, a method where AI agents iteratively improve the prompts, tools, and logic used to interact with environments. The authors argue that current evaluations are flawed because they often test evolved harnesses on the same data used for optimization, risking overfitting rather than genuine design improvement. By comparing harness evolution against simpler test-time scaling baselines—such as parallel sampling and sequential refinement—the study finds that evolution does not consistently provide superior results. Furthermore, experiments demonstrate that the performance gains from harness evolution often fail to generalize to new, unseen tasks. The findings suggest that many apparent improvements stem from memorizing task-specific shortcuts rather than distilling reusable engineering principles. Ultimately, the paper calls for more rigorous evaluation protocols that use disjoint search and testing sets to accurately measure the utility of automated agent scaffolds.

This paper studies how post-training pipelines transform large language models into effective reasoners through compositional generalization. The authors propose a hierarchical latent selection model that separates reasoning into atomic skills, such as local operations, and routing mechanisms that dictate how information is composed. Their theory suggests that supervised fine-tuning (SFT) provides the necessary raw materials, while reinforcement learning (RL) identifies and decomposes these elements into reusable modules. Controlled experiments validate that RL enables models to solve novel tasks by recombining learned atoms in ways not seen during training. Ultimately, the study concludes that SFT should focus on broad module coverage while RL should target genuinely new compositions to maximize out-of-distribution performance.

This research paper advocates for actionable interpretability as the primary standard for evaluating how effectively we explain deep learning models. The authors argue that current studies often lack real-world impact because they prioritize theoretical understanding over practical utility and concrete decision-making. To bridge this gap, the text introduces a framework and checklist designed to help researchers move beyond exploratory insights toward measurable interventions. By focusing on five key domains—including surgical interventions and alignment—the paper suggests that interpretability can lead to tangible improvements in model safety and performance. Ultimately, the work calls for a shift in academic incentives to reward findings that enable specific actions by developers and policymakers.

This paper introduces a new algorithm called first-order rejection sampling (FORS) to achieve high-accuracy sampling for diffusion models and log-concave distributions. By utilizing only score estimates (the gradient of the log-density) rather than density evaluations, the researchers provide a method that converges exponentially fast, requiring only polylogarithmic steps relative to the target error. This represents an exponential improvement over previous sampling techniques that typically scaled polynomially. The authors demonstrate that their approach is robust under minimal data assumptions, with complexity primarily determined by the intrinsic dimension of the data. Furthermore, the framework successfully addresses the log-concave sampling problem, matching state-of-the-art performance without needing complex density-based filters.

his research paper introduces a novel statistical framework for conducting causal inference using video features as treatments, a significant advancement for analyzing high-dimensional, unstructured data. To overcome the challenges of latent and dynamic confounding, the authors utilize deep generative artificial intelligence to extract low-dimensional internal representations that serve as summaries of video content. They propose a consistent and asymptotically normal estimator based on a longitudinal neural network architecture, allowing for the identification of potential-outcome trajectories under dynamic stochastic interventions. The methodology is empirically validated through a Super Mario Bros.™ benchmark with known ground-truth effects and an application to 2020 U.S. presidential campaign advertisements. Their findings demonstrate that increasing the appearance of a candidate in a video segment directly correlates with higher viewer evaluations, providing a robust tool for future social science research.

This research paper investigates the underlying mechanisms of Thompson Sampling, a popular bandit algorithm, by reframing it as an online optimization process. While traditionally viewed as a simple heuristic, the authors prove that Thompson Sampling actually minimizes instantaneous squared regret regularized by a specific measure of residual uncertainty. By comparing this mechanism to a Bellman-optimal benchmark, the study identifies a performance gap caused by Thompson Sampling's failure to account for the "tension" between exploration and exploitation. To address this, the authors propose a principled fix that adaptively shuts down exploration when the leading arm also provides the most information. Ultimately, this framework provides a theoretical compass for improving randomized algorithms by treating policy design as regularizer engineering.

This paper introduces **Off-GLADIUS**, a novel algorithm designed for **offline reinforcement learning** that utilizes **Bellman Residual Minimization (BRM)**. While traditional BRM methods often struggle with stability and convergence issues, this research proves that the proposed approach achieves **global optimality** by satisfying a **Polyak–Łojasiewicz (PL) condition**. The authors establish that for linear and sufficiently wide **neural networks**, the algorithm converges linearly to the global optimum despite the non-convex nature of the objective function. This theoretical breakthrough addresses a long-standing open question regarding the convergence guarantees of gradient-based BRM in offline settings. Empirically, the study demonstrates that **Off-GLADIUS** matches or exceeds the performance of established baselines like **Conservative Q-Learning (CQL)** and **OptiDICE** across various control benchmarks. Ultimately, the paper bridges the gap between theoretical stability and practical effectiveness, offering a rigorous framework for learning optimal policies from fixed datasets.

Researchers from Stanford, UC Berkeley, and NVIDIA have introduced LLM-as-a-Verifier, a novel framework designed to improve how artificial intelligence evaluates its own work. Unlike traditional methods that use simple pass-fail scores, this system calculates continuous scores by analyzing the underlying probability of specific words within a language model’s output. This approach allows the system to scale its accuracy by increasing score detail, performing multiple evaluations, and breaking complex tasks into simpler parts. The framework has set new records for accuracy in specialized fields like computer programming, robotic control, and medical tasks. Beyond grading results, the technology can track an agent's real-time progress and provide the detailed feedback necessary to train robots more efficiently. Ultimately, the study suggests that refining how models verify information is a critical new path for making autonomous systems more reliable and capable.

This research paper investigates language model capacity by introducing a new method to measure how much a model truly memorizes versus what it generalizes. The authors distinguish between unintended memorization, which is specific data storage, and generalization, which is the understanding of broader patterns. By testing the GPT family, they determine these models possess a storage capacity of approximately 3.6 bits-per-parameter. The study reveals that the double descent phenomenon occurs specifically when a dataset's size surpasses the model's total bit capacity. Furthermore, the researchers established scaling laws to predict the success of membership inference attacks, which identify if a specific datapoint was used in training. Their findings suggest that modern models are trained on so much data that standard membership inference is increasingly difficult for average samples.

This research paper argues that current methods for Uncertainty Quantification (UQ) in large language models are fundamentally flawed because they function as unsupervised clustering rather than measures of factual accuracy. The authors contend that these techniques merely track internal consistency, which fails to identify confident hallucinations where a model is consistently wrong. This reliance on internal stability creates a false sense of security and suffers from issues like hyperparameter sensitivity and a lack of objective ground truth. To fix these problems, the paper proposes a paradigm shift that anchors model confidence in external reality and objective verification. Ultimately, the researchers provide a roadmap for the community to develop more reliable metrics for ensuring AI safety in high-stakes environments.

This position paper discusess Theory of Agent (ToA), a framework that redefines large language model agents as decision-makers who must choose between internal reasoning and external tool use. The authors argue that agents should only invoke external tools when epistemically necessary, meaning the task cannot be reliably solved using the model's existing internal knowledge and logic. This perspective addresses common failures like overthinking and overacting, which occur when an agent's internal solvability estimates are poorly calibrated. By treating reasoning and acting as co-equal methods for reducing uncertainty, the framework highlights that unnecessary delegation to tools can stagnate the growth of an agent's internal intelligence. Ultimately, the research suggests that alignment should be measured by how effectively an agent allocates epistemic effort rather than just achieving a correct answer. These principles offer a new trajectory for training and evaluating agents to ensure they become more autonomous and efficient over time.

This research paper explores the development of efficient recommendation systems, such as AI shopping assistants, that manage multi-round interactions between a platform, advertisers, and users. The authors address a fundamental challenge: advertisers possess private, multi-dimensional information about both their own profit values and the user's preferences, creating incentives to manipulate recommendations. To solve this, the study introduces data-driven dynamic team mechanisms that align these conflicting incentives by conditioning advertiser payments on real-time user feedback. By utilizing behavioral signals like purchases and follow-up queries, the platform can create unbiased estimators of user tastes to ensure the most socially beneficial products are suggested. The proposed framework guarantees that advertisers act truthfully while maintaining individual participation and budget surplus for the platform. Ultimately, the paper demonstrates how the conversational nature of generative AI provides a unique stream of data that overcomes traditional economic barriers to efficiency in digital marketplaces.

This paper explores a surprising structural property of large language models: most reinforcement learning (RL) gains are concentrated in a very small subset of transformer layers. By isolating and training individual layers, researchers discovered that optimizing just a single middle layer can match or even exceed the performance of full-parameter RL training. This phenomenon was remarkably consistent across multiple model families like Qwen3 and Qwen2.5, various RL algorithms, and diverse tasks including mathematics, coding, and agentic decision-making. The study reveals that layers near the input and output ends contribute significantly less to post-training improvements than those in the 40%–60% depth range. Leveraging these insights, the authors developed layer-aware training strategies that prioritize these high-contribution layers to outperform standard uniform training methods. Additionally, the findings suggest that different layers capture complementary problem-solving behaviors, which can be combined through majority voting for further accuracy gains. Overall, the work challenges the assumption that RL adaptation must be distributed throughout a network and offers a more efficient, targeted approach to LLM post-training.

This research investigates the effectiveness of integrating reinforcement learning (RL) earlier in the large language model training pipeline rather than treating it solely as a final post-training step. The authors demonstrate that RL is effective remarkably early, often matching the performance of standard sequential pipelines after only a small fraction of pre-training is complete. Unlike supervised fine-tuning (SFT), which tends to degrade a model's general capabilities and narrow its output, direct RL preserves general skills and expands the diversity of reasoning paths. The study also identifies that targeted data composition is more critical for RL success than simply increasing model size. Finally, the researchers propose a parallel averaging method that combines RL and SFT updates to achieve superior results across all training stages. Together, these findings suggest that the current standard of isolating RL to the end of training is an unnecessary design choice that limits model potential.

This paper introduces a theoretical framework for language generation in the limit, exploring how machines can learn to produce valid, unseen strings from a target language through various forms of feedback. The authors specifically investigate two models: mistake feedback, where a generator learns if its prior output was incorrect, and query feedback, where the generator can actively ask if specific strings belong to the target language. A central contribution of the research is the identification of countable inner-covers as the definitive combinatorial property that determines whether a collection of languages can be successfully generated under these feedback conditions. The study proves that while access to feedback makes generation more robust to noise and contamination, it also reveals a structural divergence between element-based and set-based generators in certain query scenarios. Furthermore, the findings demonstrate that with feedback, a generator can succeed even without receiving positive examples from an adversary, relying solely on the feedback channel. These results offer new insights into the closure properties of language collections and provide a clearer mathematical foundation for understanding the mechanisms behind large language models and human learning.

This research paper explores theoretical AI alignment through the lens of Bayesian persuasion, specifically examining how a misaligned AI agent might manipulate information. The authors utilize a bit-string model to analyze the interaction between an AI sender aiming to maximize "1" guesses and a human receiver seeking accuracy. A primary contribution is the establishment of a universal upper bound, proving that the receiver's utility under a strategic AI is at most 1.5 times the utility they would obtain without any signals. The study further demonstrates that this bound becomes tighter when the information follows independent product priors, as these limit the sender's ability to exploit correlations. Conversely, the authors provide a six-bit prior example to show that specific dependencies can drive the utility ratio above 1.25, proving there are limits to how much the bound can be lowered. Ultimately, this work provides mathematical guarantees on how much useful information can still reach a human even when the AI's incentives are not perfectly aligned.

The Spiral framework addresses a limitation in current language model training where models are optimized for single-trace reasoning but fail to coordinate complex inference strategies at test time. To solve this, researchers combine set reinforcement learning with standard reinforcement learning to train models on sequential, parallel, and aggregative compute primitives simultaneously. The model learns to generate a diverse set of parallel search traces that are specifically designed to be synthesized by a downstream aggregator into a correct final response. By optimizing the entire pipeline end-to-end, the system moves beyond rigid, hand-designed scaffolds toward learned search procedures. Experimental results demonstrate that this method significantly improves scaling efficiency and performance on difficult mathematical reasoning tasks. Ultimately, Spiral enables models to effectively utilize larger token budgets through recursive self-aggregation and more sophisticated verification behaviors.

We discuss Qwen-AgentWorld, a pioneering suite of language world models designed to simulate complex digital environments for artificial intelligence agents. By training on over 10 million trajectories across seven domains, including operating systems, web browsers, and software engineering sandboxes, these models learn to predict how an environment will respond to specific actions. This simulation capability allows agents to rehearse scenarios, refine their decision-making, and learn from a vast scale of diverse interactions without needing constant access to live, physical systems. The research details a three-stage training pipeline consisting of continual pre-training, supervised fine-tuning, and reinforcement learning to ensure high fidelity in these virtual environments. Furthermore, the paper presents AgentWorldBench, a rigorous new benchmark used to verify that these world models can accurately mimic real-world dynamics. Ultimately, the authors demonstrate that integrating world modeling into agent frameworks significantly boosts performance by providing a foundation for predictive reasoning and planning.

This paper discusses a statistical framework for offline reinforcement learning using trajectory-level supervision, where only final outcomes or preferences are observed rather than step-by-step rewards. The authors introduce OPAC, a pessimistic actor-critic algorithm designed to learn from these aggregated signals by estimating latent rewards and applying pessimism to account for distribution shifts. Their analysis establishes that moving from process-level to outcome-level feedback incurs a quantifiable statistical cost, specifically an additional horizon factor in sample complexity. The research also explores generalized RL objectives, proving that non-linear outcomes like "all-success" criteria can lead to exponentially difficult learning problems. To address this, they identify specific structural coefficients, $\kappa_\mu(\sigma)$ and $\chi_\mu(\sigma)$, which determine when efficient learning remains possible. Ultimately, the paper provides a theoretical boundary for when sparse, trajectory-based data can successfully guide sequential decision-making.

SuperThoughts is a novel framework designed to accelerate the Chain-of-Thought (CoT) reasoning process in large language models by processing tokens in superposition. Unlike traditional models that generate tokens sequentially, this method uses a compressor to fuse pairs of consecutive tokens into single latent representations, effectively halving the number of required forward passes. To ensure accuracy is not sacrificed for speed, the system employs a Multi-Token Prediction (MTP) module and a confidence-based adaptive mechanism that reverts to standard decoding when the model is uncertain. Experimental results on complex mathematical and scientific benchmarks show that SuperThoughts reduces reasoning length by 20–35% while maintaining performance within a few percentage points of the original baseline. The research highlights that larger models are particularly adept at handling this compression, achieving significant wall-clock time reductions during inference. Ultimately, this approach offers a more efficient way to utilize test-time compute without losing the dense supervision provided by discrete token training.

This research paper introduces First-Explore Proximal Policy Optimization (FE-PPO), a new reinforcement learning algorithm designed to improve how agents discover rewards in complex, deceptive environments. While standard meta-learning methods often fail when immediate rewards are misleading, the FE-PPO framework trains agents specifically to gather information during exploration that will maximize success in later exploitation phases. By integrating a value function and bootstrapping into the original First-Explore objective, the authors significantly increase efficiency, achieving high performance with 10 to 40 times fewer samples. The study demonstrates that FE-PPO consistently outperforms the strong RL² baseline across various challenging benchmarks, including navigation tasks and bandit problems. Additionally, the authors provide a more competitive comparison by implementing a Transformer-XL architecture for their baselines. Ultimately, this work offers a practical, open-source foundation for future research into efficient meta-exploration strategies.

This research paper investigates self-distillation as a powerful regularization technique for pretraining language models when high-quality data is in short supply. By comparing various training strategies across different model scales and data scarcity levels, the authors demonstrate that self-distillation significantly outperforms both direct training and standard methods like weight decay or exponential moving averages. The study identifies a specific crossover threshold where distillation becomes superior, particularly when the available data is less than one-fourth of the amount prescribed by Chinchilla scaling laws. Practical results suggest that using larger models with natural teacher temperatures provides the most effective supervision, preventing the rapid overfitting typically seen in data-constrained environments. Ultimately, the work advocates for self-distillation as a robust alternative for improving model performance when compute resources outpace the available data pool.

eta-Harness is an advanced optimization system designed to improve how language-model agents process and compress long interaction histories into useful states. Unlike traditional methods that rely on manual engineering or simple feedback, this system uses a coding agent to search for and rewrite the "harness" code that manages an agent's memory and retrieval. By providing the proposer with direct filesystem access to raw execution traces and historical performance data, it avoids the information loss associated with summarized feedback. This approach allows the system to discover superior strategies for history summarization and adaptive retrieval across various complex tasks. Experimental results demonstrate that Meta-Harness achieves top-tier performance on benchmarks like TerminalBench-2 and improves accuracy in mathematical reasoning and text classification. Ultimately, the research suggests that the way agents construct their own internal state can be optimized as an embedded learning problem.

Exploratory RL (ExpRL) is an automated mid-training method designed to enhance the reasoning capabilities of large language models before they undergo standard reinforcement learning. While traditional reinforcement learning often struggles with sparse rewards on difficult problems, ExpRL uses human-written reference solutions as reward scaffolds to provide dense, informative feedback on partial progress. This approach employs an LLM judge to evaluate on-policy reasoning traces against specific rubrics, assigning rewards at both the outcome and process levels to reinforce productive intermediate steps. By shifting probability mass toward successful solution strategies, the method significantly improves pass@k performance and broadens the model’s coverage of complex reasoning paths. Experimental results demonstrate that ExpRL creates a superior initialization for subsequent training, outperforming supervised fine-tuning and standard distillation across challenging math and science benchmarks. Ultimately, this technique fosters sophisticated behaviors like self-correction and backtracking, which are essential for solving high-level reasoning tasks.

This paper introduces a statistical framework for making valid scientific discoveries using synthetic data, specifically addressing concerns that artificially generated data can be biased or noisy. The authors propose a new technical condition called task exchangeability, which allows researchers to calibrate synthetic results by comparing them to historical tasks where both real and synthetic data are available. By measuring the discrepancy between real and synthetic outcomes in these past cases, the method can adjust confidence intervals for new tasks where only synthetic data exists. The researchers demonstrate that this approach provides provable validity guarantees across various fields, including social science surveys and AI evaluation. Experiments show that while naive synthetic-only intervals are often severely biased and overconfident, the task-exchangeability method consistently covers the true values. Ultimately, this framework enables scientists to use LLM-generated "silicon samples" and automated raters to accelerate discovery without sacrificing statistical rigor.