All papers

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

The paper evaluates whether self-distillation using privileged information improves reasoning capabilities or reflects overfitting to training tasks.

Examines operational challenges of maintaining and deploying fleets of software variants generated by large language models.

Trains an unsupervised vision-language reward model using intermediate layer representations and caption perturbations without human preference labels.

Measures representation drift across continual learning methods, finding n-gram memory updates cause less internal drift than LoRA.

This paper introduces a benchmark to diagnose the ability of video language models to count repeated action cycles.

This paper introduces a timestamp theory and persistent ledger to route updated information to pending LLM decisions before action.

The authors evaluate language models on finance peer review tasks, finding predictive ability for publication outcomes through correctness rather than originality assessments.

Studies how chain-of-thought reasoning affects social bias and code quality during LLM code generation.

Studies how accelerated decision-making models affect overall task completion, latency, and costs across embodied and mobile agents.

Proposes a two-phase token pruning strategy for vision-language models and identifies evaluation flaws in common decode-phase benchmarks.

This paper audits key-cache compression and weight pruning techniques on a small edge-deployed language model testbed.

This paper introduces a benchmark to measure how language model agents handle ambiguous or missing goals, context, and constraints.

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

The paper studies how outdated code comments systematically mislead language models during automated code completion.

Develops multi-layer linear probes on model activations to detect agent deception and sabotage across diverse scenarios.

This paper introduces an observation speculation framework using small world models to accelerate reinforcement learning rollouts for tool-using agents.

This paper analyzes how different visual representation types affect the 3D spatial reasoning capabilities of multimodal large language models.

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

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

This paper presents a user study evaluating how programmers interact with and perceive AI code generation tools in daily workflows.

Applies psychometric auditing to HELM Safety benchmarks to test whether harmful refusal functions as a single measurable attribute.

The paper introduces an interactive web-based tool for analyzing and visualizing execution trajectories of software engineering agents.

This paper introduces a graph-guided layer fusion method that identifies and merges functionally redundant non-adjacent layers to compress language models.

This paper introduces a training-free framework combining multi-hypothesis search and local visual verification to ground interface instructions in graphical user interfaces.

This paper examines how single-agent long-term memory accumulates repeated historical claims and introduces a benchmark and framework to mitigate unverified persistence.

This paper introduces a benchmark to measure how language model agents handle ambiguous or missing goals, context, and constraints.

Proposes an evaluation framework to measure how LLM agents handle uncertainty and resolve conflicting knowledge during execution.

Uses sparse autoencoders to select diverse training samples and prevent underrepresenting tail concepts across vision and multimodal models.

The paper proposes training looped transformers at variable depths only during a short finishing stage to reduce compute costs.

This paper introduces an approach and benchmark for evaluating whether language model agents can infer unstated true requirements instead of following incomplete instructions.

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

The paper cleans code smells from training datasets to evaluate and improve the quality of LLM-generated code.

Evaluates model coherence using 175 specific questions to measure internal contradictions and identity conflation under answer resampling.

The paper examines using transfer probes to group business tasks when sharing Low-Rank Adaptation experts under fixed parameter budgets.

Develops an agent-centric 3D language model that grounds spatial understanding in physical poses without requiring text situation descriptions.

This paper conducts an adversarial evaluation comparing a probability-based safety judge against large language model judges and safety classifiers across multiple attack types.

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

The paper proposes injecting reverse-engineered architectural and domain knowledge into coding agents to prevent conceptual and structural code errors.

Analyzes training data extraction risks in language models when prompts include contextual conditioning such as retrieved documents.

The paper investigates dynamic tensor rematerialization, identifying cases where increasing memory budgets causes training to unexpectedly run out of memory.

Develops an interpretable AI framework for breast ultrasound diagnosis incorporating clinical standard descriptors and assessment workflows.

This paper examines how single-agent long-term memory accumulates repeated historical claims and introduces a benchmark and framework to mitigate unverified persistence.

Presents ECMR, a dynamic evidence-aware controller that guides multi-agent deliberation steps to prevent false consensus and reduce redundant reasoning.

This paper introduces a multi-agent system that evaluates code extensibility and detects SOLID design violations through automated code extension.

The paper demonstrates that machine-generated text detectors rely heavily on document formatting and PDF extraction artifacts rather than text quality.

The paper examines teacher interventions along student generation paths to identify how teacher guidance drives performance gains in on-policy distillation.

This paper combines evolutionary computation with language models to optimize register-transfer level hardware code for correctness and efficiency.

The paper proposes a framework that evolves a reusable library of skills and tools based on experience for spatial reasoning agents.

This paper introduces an approach and benchmark for evaluating whether language model agents can infer unstated true requirements instead of following incomplete instructions.

Introduces ProtoArena, a black-box dynamic benchmark evaluating the ability of AI agents to discover and exploit cross-component network protocol vulnerabilities.

Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations.

This paper shows that routing statistics across task phases fail to provide an effective guide for mixed-precision quantization in mixture-of-experts models.

The paper introduces a method to detect adversarial chain-of-thought hijacking in backdoored multimodal reasoning models by monitoring reasoning dynamics.

Evaluates enterprise policy enforcement across structured JSON versus single-token outputs, measuring the impact of explicit reasoning on decision stability.

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

The paper presents a multi-agent framework that dynamically generates execution context to enable partial code execution.

Evaluates whether automated prompt optimizers inadvertently cause regressions in unscored prompt constraints and requirements.

This paper proposes Lookahead Self-Distillation, providing cross-instance token-level supervision to improve reinforcement learning with verifiable rewards for language models.

The paper analyzes failure modes in unified visual generators using self-reward loops and proposes stabilization methods to prevent quality degradation.