All papers

The paper identifies an activation steering vector that controls whether language models rely on internal memory or retrieved external context.

The paper introduces an algorithm to optimize denoising timestep placement in continuous diffusion language models to improve generation quality.

The paper presents a latent world action model that predicts action-induced state transitions rather than full-screen pixel reconstructions for graphical agents.

Designs diagnostic probes to determine whether audio-language evaluation models judge music based on audio or prompt content.

Introduces AgenticBBO-Bench, a benchmark across multiple domains to evaluate LLM agents on black-box optimization under unified finite budgets.

Evaluates the ability of language models to predict state changes in discrete spatial grids across sequential operations.

The paper evaluates failure modes and confidence calibration across frontier and lightweight language models during entity extraction from cyber threat intelligence reports.

The paper presents an empirical scaling law study measuring parameter and compute trade-offs in spiking neural network language models.

Introduces a benchmark evaluating streaming vision-language models on fast-moving, high-dynamic events in real-world video streams.

Proposes a recurrent harness framework that uses streaming feedback to revise textual memory for continual agent learning.

The paper proposes a training-free interface for heterogeneous language models to communicate facts using shared symbolic word indices.

The paper introduces a distillation method that trains agent policies on explicit runtime control semantics recovered from replayed execution traces.

This paper presents a lightweight, interpretable neural network architecture for resource-constrained time-series classification on microcontrollers.

This paper presents a multimodal fusion network for depression detection combining gait, audio-visual data, and textual personality descriptions.

The paper trains 123 models across 54 tokenizers to show tokenizer choice matters more for languages with less training data.

The paper introduces a router-free method that adaptively thresholds rank-one LoRA adapter responses to remove cross-talk noise per token.

Evaluates patient linkage and attribute inference risks present in brain magnetic resonance imaging representations generated by pretrained foundation models.

This paper presents a training framework that teaches language models to orchestrate tools through executable code instead of sequential prompt accumulation.

Presents a multimodal benchmark testing whether models can explain observed chemical effects by linking structural features to reaction mechanisms.

The paper develops an approach to map AI semantic blind spots and evaluates the quality of distinct human ideas.

Introduces a dynamic precision control framework that allocates weight chunks to mixture-of-experts models during inference.

The paper evaluates an artificial intelligence decision model across cognitive reflection, classical bias benchmarks, and pricing tasks.

The paper investigates an asynchronous training recipe combining the Muon optimizer, token-averaged loss, and penalties for long-horizon agent tasks.

The paper uses attention maps from vision-language models to directly guide pixel decoders for referring image segmentation.

Measures post-training activation geometry shifts against run-to-run noise, identifying distinct structural signatures left by supervised fine-tuning.

Presents a large graphical user interface grounding dataset and a combined distillation and reinforcement learning training method for agents.

The paper presents a red-teaming method that injects stealthy attack operations across interdependent agent skills using code rewriting.

Introduces EmoGenie, a benchmark evaluating expressed and evoked emotions in text-to-image generation and multimodal language model understanding.

The paper proposes a query router that balances answer utility, security risks, and token budgets when selecting language models for mixed enterprise workloads.

This paper applies recursive training and discrete deduction procedures to message-passing neural networks to solve constraint-reasoning tasks with few parameters.

The paper introduces an evaluation framework to diagnose step-level execution stability and error recovery in graphical user interface agents.

Proposes LongForcing, a temporal manifold matching distillation approach that scales long-horizon video rollout stability.

The paper presents AutoEnvScaling, using terminal agents to automatically generate reinforcement learning environments from model rollouts.

This paper presents an interpretability protocol to monitor latent reasoning dynamics and stability across multiple looped language model families.

Proposes a spectral representation framework to attribute agent failures to specific decisive steps using only trajectory-level outcome labels.

The authors present a market-based post-training framework that allows decentralized agent swarms to learn both action policies and coordination mechanisms without central orchestration.

The paper compares typed probabilistic decision models with large language models on tabular data classification and format robustness.

Benchmarks multimodal large language models paired with object detectors for zero-shot human-object interaction recognition.

The paper evaluates the energy efficiency, carbon intensity, and accuracy of ten post-training quantization methods across various language models and GPUs.

The paper presents a framework and benchmark to evaluate whether self-evolving robotic agents genuinely learn reusable skills across tasks.

The paper presents a method that uses agent failure patterns to generate tailored simulation environments for embodied navigation learning.

This paper analyzes three distinct failure modes where reinforcement learning agents misgeneralize goals despite competent behavior.

Proposes FlowCred, a hierarchical credit assignment framework to coordinate prompt optimization and video generator post-training.

The paper demonstrates multi-fidelity training to optimize data generation compute and error scaling for neural PDE surrogates.

Shows that activation steering evaluations can conflate bias reduction with model abstention, introducing a gradient-constrained direction design to isolate preferences.

The paper analyzes five internal representation reading methods to examine how they detect prompt injections in language models.

This paper introduces a closed-loop world-action model that improves generalist robot policies using deployment rollouts of mixed quality.

The paper presents a simulated benchmark across eleven unconstrained environments to evaluate vision-language models on documenting real-world human physical work.

The paper compares single-head autoregressive beam search with parallel multi-token prediction for generating semantic IDs in recommendation systems.

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

The paper proposes a self-supervised method that uses repetitive tool calls as negative signals to reduce looping behavior in long-horizon agents.

This paper introduces a benchmark evaluating high-level reasoning and multi-step planning in vision-language models for embodied tasks.

The paper studies expert width and dimension balance in latent mixture-of-experts architectures to optimize parameter capacity across model layers.

The paper introduces an evaluation method that separates question answerability detection from answer span extraction in radiology report question answering.

Introduces Cowork-Gym, a framework synthesizing multi-tool collaborative work environments and user simulations to train enterprise agents.

The paper introduces a self-training method that uses runtime policy verification feedback as rewards to teach agents safety compliance.

This paper introduces a benchmark to analyze whether visual reasoning skills evaluated in toy games transfer to real-world tasks.

RSIPO introduces a reinforcement learning framework where agentic skills and policy models co-evolve using self-extracted trajectory experience.

This paper updates a long-horizon terminal benchmark to measure sustained agent progress across complex engineering workflows using dense process rewards.

The paper presents an in-context learning method that extracts compact concept vectors from demonstrations to guide sequential action choices for agents.