LLM agents face challenges in real-time, dynamic environments where reasoning and tool calls consume time. RealtimeBench is introduced to evaluate agents while games run concurrently, showing performance differences between models under these conditions. Two-Timescale Offline Recursive Self-Improvement (TORSI) is proposed to improve agents by having a slow teacher guide learning while a fast student acts during gameplay. This method improved Jev's test scores significantly across multiple game types. The research advocates for real-time intelligence to keep LLM agents effective in continuously changing worlds.
Benchmark saturation occurs when multiple models achieve near-perfect scores, diminishing their utility for differentiating performance. The study investigates whether the detailed transcripts of how models solve tasks retain capability information even after benchmark scores have saturated. Researchers tested methods like pairwise judging and feature extraction from transcripts to assess these retained capabilities. Pairwise judging proved most effective, reducing prediction error significantly compared to using only benchmark scores. Saturated benchmark transcripts contain capability-predicting signals, but robust extraction methods are needed for extended benchmarking.
Reward-guided discrete diffusion involves computationally expensive Sequential Monte Carlo (SMC) guidance at every denoising step. VISTA is a training-free method that automatically selects guided timesteps under a fixed budget by approximating the full-guidance path. It uses value statistics from warm-up generations to optimize the selection via dynamic programming. VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules across various generation tasks. This method significantly reduces per-generation latency while maintaining or exceeding full-guidance performance.
AlphaDiana is a measurement system designed for agents engaged in verifiable reasoning tasks. It provides modular interfaces for datasets, agents, environments, and scorers, while preserving harness-native control loops. The system allows for macro-analysis of complete agents and micro-interventions on components like tools, skills, and memory under matched conditions. Evaluation shows that harnesses and component interventions do not consistently improve reasoning accuracy alone. AlphaDiana links answer correctness and execution status to traces and resource records for agent comparison and failure diagnosis.
Correctness feedback can suppress the evidence needed for specification-completeness auditing, which aims to find violations of intended behavior. Experiments on Verus/Rust tasks showed that independent no-feedback discovery yielded more certified witnesses than the corrective-feedback loop. Phase-separated execution methods successfully matched aggregate yields without reusing responses, leading to stronger specification repair. Preserving witnesses resulted in higher implementation success rates for contracts compared to those using corrective feedback. This research suggests that completeness auditing and correctness repair have opposing immediate goals.
Counterfactual Neighborhood Risk Fields (CNRF) is an evidence-grounded framework for selectively ensuring safety in Vision-Language Models (VLMs). It constructs context-matched benign-harmful counterfactual pairs to derive capability-related risk directions from representation differences. CNRF separates neighborhood support from risk to prevent confusing unfamiliarity with actual harmfulness. This allows for selective intervention based on victim-adaptive interfaces that route requests for review. Experiments show CNRF achieves high macro defense success rates while maintaining utility for benign tasks.
IneqAgent is an agentic framework that combines LLM reasoning with symbolic computation to prove mathematical inequalities in Lean4. It utilizes Lemma-level Monte Carlo Tree Search for coordinating lemma construction, goal reformulation, and symbolic backend selection. Backends like sum-of-squares and successive difference-substitution generate exact certificates for local proof generation. The agent integrates these proofs with LLM steps and uses Lean feedback to construct complete, verified proofs. Benchmarks show IneqAgent achieved a high success rate in solving inequality problems while demonstrating efficiency gains over direct agent methods.
L2G-SR is a framework for LLM-based symbolic regression that uses a learned structural verifier to guide the search process. It generates a structural supervision signal using tree edit distance, which accounts for commutative operators and alternative expression forms. The verifier first aligns numerical observations with formulas and then scores the structural similarity of candidate expressions to the target. At search time, the verifier aids in selecting mathematical expressions by complementing numerical error. This approach demonstrates the value of learned structural guidance for improving LLM-based symbolic regression accuracy.
The paper explores how quickly reality can change relative to an intelligent system's ability to perceive, reason, and act reliably. It introduces the Speed of Coherent Cognition ($S_C^{\mathrm{eff}}$) as a new benchmark to measure reliable AI scaling across a growing observation-reality gap. The challenge involves systems inferring reliable actions for future states as the reality gap expands. Reinforcement-learning agents were tested to see how they handled increases in reality speed while maintaining behavioral reliability. The architecture $Co^4$, inspired by cellular mechanisms, demonstrated superior performance in achieving this coherent cognition benchmark compared to Transformer and gated-MLP models.
A framework called ARIA is introduced for learning how to repair ongoing multi-agent executions in LLM-based systems. The method learns when and how to intervene during execution to correct local mistakes before they cause task failure. This is achieved by constructing a large dataset of execution-verified repairs from various multi-agent scenarios. The ARIA framework uses these repairs to train a compact model to recognize intervention points and apply effective correction strategies in runtime. The resulting model demonstrates strong repair performance across different agent frameworks and host models.
FlowBudget is a training-free method to estimate the sampling budget for flow matching models, addressing the dependence of inference cost on numerical integration. It uses simulation-free interpolation probes of the learned dynamics to estimate the leading one-step discretization error. FLOWBUDGET normalizes this error by motion and maps a numerical tolerance to a step budget based on the solver order. The method derives estimators for numerical solvers like Euler and Heun for ODE and SDE sampling. Experiments on various flow-model architectures validate FLOWBUDGET's ability to estimate the Necessary Function Evaluations without requiring multiple sample generations.
The method expands the exponential attention kernel into a truncated Taylor series of $n$ terms. This factorization collapses causal history into a fixed number of recurrent states, which are updated by rank-one outer products independent of context length. This recursion is lossless, and the states are additive across chunks, allowing for parallel prefill scans. The resulting decoding latency remains nearly constant per token, offering significant speedups as context length increases. Optimal performance is achieved with $n=3$ terms, providing a good balance between accuracy and cost.
Research investigates using feature elicitation for Sparse Autoencoder (SAE) steering to mitigate deceptive alignment behavior in small Large Language Models (LLMs). The study tests whether steering SAE features associated with deontological ethics can reduce the compliance gap observed in alignment faking. A novel bag-of-words pipeline was developed to identify concept-level steering directions for this purpose. The intervention significantly reduced the compliance gap in Gemma-3-27B-IT but was ineffective in LLaMA-3-8B-Instruct. Deceptive defanging, where harmfulness is degraded while helpful compliance is maintained, was also identified as a form of deceptive behavior.
Reinforcement learning with verifiable rewards (RLVR) can yield domain-specific gains, such as higher accuracy in mathematics, which are often extracted via answer format. The study investigates how much of this out-of-domain gain stems from the answer format rather than the answer content itself. Experiments on 7B SimpleRL pairs show that 82% of the reported out-of-domain multiple-choice gain disappears when considering the format. This format gain occurs because the reinforcement learning process moves answers into a box that extractors prioritize. The magnitude of this format gain is dependent on the specific answer extractor used to read the output.
KNOU-Flood advances global streamflow forecasting using lead-aware probabilistic machine learning. The method extends an open recipe by incorporating a longer input window, lead-aware forecast inputs, GraphCast, and a latent-noise ensemble. Results show KNOU-Flood achieves superior performance in streamflow forecasting compared to existing methods, particularly for longer lead times. The probabilistic head significantly contributes to the performance gains observed. The approach also effectively detects larger flood events and improves calibration with inference-time dropout.
The study investigates what spectral summaries and learning curves reveal about centered quadratic networks. The work explores how factorized loss curves differ based on cycle invariants, providing bounds on the required observations for identification. Specific conditions, such as Gaussian inputs and rationally independent teacher spectra, allow for exact scalar loss curve identification of the initial predictor and the unknown teacher spectrum. Finite exact global bounds exist for known teachers, and exact certificates can verify the constructions, although numerical inverse checks reveal ill-conditioning.
Sparse autoencoders require less data for functional behavior than for feature reproducibility. Functional reconstruction often approaches the full-data reference at smaller data budgets, while feature reproducibility improves more gradually with more data. Data selection methods affect performance, with random sampling favoring reconstruction and diversity-based sampling improving AutoInterp. Theoretical analysis suggests that feature recovery requires stronger identifiability and optimization conditions than mere statistical consistency. Explicit reproducibility checks are necessary to ensure feature-level claims in SAE research.
TurboMamba introduces a language-model architecture designed for faster inference, training, and improved token prediction using lexical memory. It employs a global Mamba-3 model running per token block, complemented by a small Transformer for local predictions. Lightweight additions optimize local predictions through direct context access, score adjustments, and frequency mixtures. Comparisons on FineWeb-Edu show TurboMamba achieves lower prediction loss and significantly faster training throughput and reduced batch latency than baselines like Mamba-3. Ultimately, it demonstrates substantial speedups over other models while maintaining superior prediction quality.
The study investigates if the habit of writing Python when not explicitly told a language is due to a small set of components within CodeLMs. Researchers analyze this habit across various model-language settings, prompt suites, and component budgets. Findings indicate that the preference is large, grows with model size, and is influenced by the probing method used. The perceived localization of this preference depends on the budget for component search, with larger budgets revealing more localized circuits. Interventions like patching and steering can successfully influence the target syntax, demonstrating the importance of analysis choices.
A framework is proposed to learn topology-aware representations for mechanical Triply Periodic Minimal Surface (TPMS) metamaterials. This method integrates persistent homology into a DMAE backbone to explicitly encode global structural connectivity relevant to mechanical response. A persistence-guided masking strategy and an auxiliary decoder are used to ensure the latent representation retains information about topological features across connected components and voids. Evaluation on mechanical and family classification tasks shows improved performance, particularly in unsupervised property-clustering, establishing persistent topology as a useful inductive bias. The results demonstrate that topological persistence is effective for learning complex 3D mechanical metamaterial representations.
A polylogarithmic bound for list replicability is presented in the paper. This work provides a short, self-contained proof for a bound on this quantity. The bound is established for the list-replicability number of every class of $M\ge 2$ binary concepts on $N\ge 2$ domain points. The established bound is $O(\log^2 M\log N)$. This result offers a polylogarithmic upper bound for list replicability.
Latent Annealed Posterior Sampling (LAPS) is proposed to address challenges in inferring random fields from sparse observations by reducing dimensionality and learning a latent prior using flow matching. LAPS samples the posterior entirely in the latent space using an annealing procedure that incorporates observation likelihood. A variance-preserving bridge is introduced to control covariance inflation during sampling. Experiments show LAPS accurately approximates posteriors, preserves modes, and reduces sampling cost dependence on spatial resolution.
Graph Neural Additive Networks (GNANs) have limitations in disentangling features, topology, and interactions due to aggregated relevance scores and distance-based representations. AMIGA is introduced to address this by representing topology via mined motifs and decomposing node contributions into feature, structure, and feature-structure interaction channels. This decomposition centers the channels around a graph-balanced baseline, ensuring a well-defined and unique separation of effects. AMIGA outperforms GNANs on ten benchmarks while providing node-local explanations. Consequently, AMIGA offers interpretable insights into graph representations at various granularities.
Flow maps utilize stop gradient operators and adaptive weighting, yet the impact of these heuristics on training dynamics remains unanalyzed. The work provides a grounded theory explaining the necessity of these methods by analyzing the functional gradient flow from MeanFlow and Lagrangian losses. The analysis confirms that these objectives converge to the true flow map, establishing their validity. Adaptive normalization provides bounded updates, and with specific weighting, offers convergence guarantees to a solution neighborhood. Experiments show that proper reweighting is essential for strong empirical performance on image datasets.
The study investigates identifiability in supervised learning by examining when neural networks can recover true latent variables from data. Two data generating processes are considered: generative, where labels cause latents, and discriminative, where latents cause labels. Identifiability fails when the representation has insufficient dimensions, leading to nonlinear encoders, but constraints can allow for recovery of individual latent variables. In the generative setting, certain constraints imply each neuron encodes exactly one latent, while in the discriminative setting, sparse latents can be perfectly recovered under specific conditions. This work unifies concepts from nonlinear ICA and compressed sensing to delineate the transition between nonlinear distortions and linear representations in neural networks.
The Time-partitioning Gaussian Process Mixture Model (TPGPMM) is a Bayesian nonparametric framework designed to identify persistent market regimes and model nonstationary time series dependence. It integrates an adaptive Pitman--Yor prior with regime-specific Gaussian processes, avoiding strict Markovian dependence within regimes. A split--merge Metropolis-within-Gibbs sampler is used to jointly learn change points and model parameters, quantifying uncertainty over segmentations. Simulations show the model effectively recovers the regime structure, successfully identifying market events in data like the VIX. The model demonstrates a trade-off between interpretable segmentation and predictive accuracy, with variants adjusting observation noise for improved prediction.
The work introduces Proximal Survival Bridge learning (PSB), a flexible doubly robust proximal framework for censored outcomes. It aims to address unmeasured confounding in causal inference by leveraging proxy variables beyond restrictive parametric models. The framework yields outcome-bridge, treatment-bridge, and doubly robust estimators using two flexible learners: PSB-K (kernel estimator) and PSB-N (neural estimator). Experimental results show PSB-K performs best on SEM and ACTG benchmarks, while PSB-N is most stable on the Framingham study.
The paper investigates the statistical behavior of Stochastic Gradient Descent (SGD) with a constant step size under iterated random functions. It establishes convergence in Euclidean distance by demonstrating the Geometric Moment Contraction (GMC) of SGD. The work proves quenched central limit theorem, invariance principle, and $\alpha$-stable convergence for averaged SGD, irrespective of initialization. The authors analyze the impact of step sizes and propose methods to improve ASGD estimates closer to the global optimum. Finally, a new online inference method and a bias-reduced variant for extrapolated ASGD are introduced.
Deep learning models using ensemble spread for streamflow forecasting face issues at extreme flow conditions. Prediction error is significantly larger than the spread at high flows, increasing by a factor of up to four. Under-dispersion, where the model underestimates uncertainty, is severe in extreme flows, reaching an $18\times$ discrepancy. While ensemble spread ranking remains near-oracle for uncertainty, deployment relies on this signal despite systematic underestimation of risk for critical extreme events. Consequently, the uncertainty signal in ensemble spread may systematically underestimate the true risk during flow extremes.
A statistical framework is developed for diffusion transfer by analyzing a shared pooled score alongside a target-specific offset. The framework derives convergence rates by separating the costs associated with learning these two components. It identifies when a simpler offset leads to a faster risk-bound rate than separate fitting. The results show that the offset remains bounded and jointly regular as the diffusion time approaches zero. This work provides total-variation guarantees for diffusion models when viewed as density estimators.