All papers

Introduces Sauron, an evidence-first agentic verifier that checks both environment state and execution traces against original sources to validate task rubrics.

This paper proposes a training framework that triggers supervised fine-tuning specifically on failure states during reinforcement learning to prevent agent forgetting.

This paper presents a compact decision model designed to screen malicious agent tool skills with high recall directly from specification files.

This paper proposes a credit-assignment method that uses rubric evaluation outcomes to allocate token-level advantages during reinforcement learning.

This paper proposes an end-to-end framework integrating nonlinear dimension reduction with longitudinal counterfactual prediction for high-dimensional observational data.

This paper provides four auditing tests to determine whether failure-attribution judges actually perform multi-step checks or merely rely on spurious surface cues.

This paper introduces a framework that quantifies delegation flexibility to analyze the spectrum between centralized and decentralized multi-agent coordination.

This paper proposes Gaussian process bandit algorithms with selective verification audits to optimize AI system configurations under noisy evaluations.

This paper introduces a framework combining language model agents with deterministic software controls for reliable sensor-based industrial fault diagnosis.

This paper proposes a flow-matching method that averages Bellman target directions under a posterior distribution to improve reinforcement learning critics.

This paper introduces a benchmark to evaluate whether forecasting models can incorporate forward-looking contextual information into accurate time-series predictions.

This paper introduces a framework that enables language models to optimize physical structural designs through simulated fracture experience and hypothesis revision.

This survey analyzes how verifier noise affects reinforcement learning with verifiable rewards across 58 studies and summarizes practical mitigation techniques.

Introduces Agentic-MatriBO, a modular Bayesian optimization framework using LLM agents and deterministic compilation for material and chemical experiment design.

Provides a tutorial analyzing theoretical expressiveness gains and parameter tradeoffs when re-uploading data in quantum amplitude embedding architectures.

Introduces FASHION-RQ, an automated framework deriving hierarchical evaluation benchmarks for fashion retrieval directly from catalog structures.

Releases MMTA, a multimodal benchmark testing the ability of language models to jointly analyze time series, text, and visual data.

Proposes TStruct, a network that captures recurring temporal relationship structures across variables to improve long-term time-series forecasting.

Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages.

Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication.

Presents Interaction-as-a-Judge, using computer-use agents to evaluate interactive web and desktop software applications through direct dynamic interaction.

Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models.

The paper proposes a shopping basket choice model using specialized negative sampling to capture product interactions from retail transactions.

The authors show that medical vision-language models frequently abandon correct clinical answers when prompted with fabricated medical documents.

The paper proposes predictability-aware supervised fine-tuning to resolve training and inference order mismatches in diffusion language models.

SHELDON introduces a dual-search reasoning framework that independently explores hypotheses and observation spaces to avoid confirmation bias.

The authors propose methods to train machine learning potentials that transfer across quantum-chemical methods to exceed single-method training accuracy.

The paper proposes a neural operator and state-space framework for high-fidelity 3D turbulent flow reconstruction from coarse data.

This paper presents an intermediate-representation method to audit whether specific private data was used in LoRA fine-tuning of language models.

The authors prove and demonstrate empirically that a two-layer attention Transformer can learn and implement in-context Q-learning algorithms.

The paper examines conceptual parallels between measurement science and machine learning uncertainty evaluation to propose a more unified assessment framework.

WorkRover introduces a framework to train long-horizon cowork agents by recursively evolving contexts, tasks, and verifiers based on execution feedback.

This paper introduces an evaluation framework grounded in published papers to assess continuous multi-stage execution in automated research agents.

This paper shows chain-of-thought monitoring can fail even when reasoning is necessary, proposing a multi-level hierarchy to measure monitorability.

This paper presents a post-hoc multicalibration framework to correct systematic biases such as position and model family in LLM evaluations.

This paper formalizes graph-structured missing features where missingness spreads across edges, testing existing graph propagation methods under this setting.

This paper defines conditions under which learned multidimensional representation blocks preserve causal abstractions and high-level concept interventions.

This paper shows data ownership determines whether agents should broadcast or relay table queries, and proposes ownership-aware routing.

This paper studies how different framings of user claims induce sycophancy in large language models on factual questions.

This paper establishes functional convergence bounds for wide neural networks approximating Gaussian processes on arbitrary domains.

This paper presents a neuro-symbolic graph transformer with first-order logic loss to recognize hierarchical human intentions from multimodal activity sequences.

TextualRL updates agent context prompts using groupwise trajectory comparisons rather than isolated feedback runs to improve policy adaptation without parameter tuning.

This paper audits a specific scalar feature scaling method to test whether claimed curvature benefits hold for domain generalization benchmarks.

Runoff eliminates option-order bias in multiple-choice model predictions by scoring items independently before running balanced candidate comparisons.

LiFUSE uses frequency-spatial self-supervised pretraining on non-contrast CT scans to improve representation learning for liver disease detection.

The authors demonstrate that encoding malicious instructions in agent skills bypasses scanners and causes models to execute harmful actions upon decoding.

JevGPT combines fast local evidence assessments with deliberative planning to improve autonomous research agents operating under fixed retrieval budgets.

The authors analyze how different optimizers leave distinct implicit bias fingerprints that influence training data reconstruction from model checkpoints.

SimInter simulates conversational recommendation users by conducting synthetic interviews between language models to enrich implicit behavioral profiles.

The authors propose a probe-based method to measure and calibrate systematic scoring biases in black-box LLM evaluation judges.

The authors trace internal representations across prompt optimization techniques to show how discrete prompts, prompt tuning, and prefixes alter model processing.

The authors reproduce and compare five language model factuality alignment methods on a standardized benchmark across identical base models and evaluation protocols.

The authors present an energy-based prior for cellular perturbation models to better distinguish specific therapeutic signals from general biological background noise.

The authors analyze readout scaling in sparse autoencoders and propose canonicalized coordinates to improve the recovery of underlying generative factors.

The authors present a training framework and dataset that enables terminal coding agents to self-improve by recovering from execution errors.

The authors identify a derivation gap where LLM judges reject correct tool arguments lacking direct textual references, and propose a dedicated reference model.

The authors propose evaluating autoformalization faithfulness by aligning phrase-level correspondences between informal mathematics and formal code under contextual interpretations.

The paper analyzes the geometric distribution of language model output logits, identifying a consistent log-linear relationship that emerges during pretraining.

The paper introduces an agent-based approach that adapts search spaces to optimize post-training recipes and configurations for pre-trained language models.

The paper analyzes implicit morals in news coverage of artificial intelligence across four global regions, identifying distinct regional variations in narrative caution.