All papers

The authors present an engine to synthesize environments, tasks, and rubrics for training and evaluating data analysis agents.

This paper examines neuron relevance characterization methods based on magnitude and distribution support for selecting instruction fine-tuning data.

The paper proposes a knowledge distillation method that transfers text language model capabilities to speech-to-speech models using internal text interfaces.

Tests delayed prompt injection attacks on life-support control agents in a simulated lunar habitat to see if safety duties persist.

Presents a simulation benchmark evaluating how embodied agents continually acquire basic skills and recombine them for complex tasks.

The paper introduces a hierarchical reinforcement learning method designed for agentic tasks that span multiple interaction sessions.

The paper evaluates whether providing page images to vision-language models improves the detection of errors in Japanese-English manga translation.

Shows that running fewer layer loops in recurrent-depth models significantly reduces safety refusals while preserving most reasoning performance.

HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding.

Introduces a benchmark evaluating multimodal language models on extracting data, tracking narrative flows, and answering questions from animated data videos.

Heterogeneous Structured Pruning selects sublayer-specific sparsity patterns guided by latency and memory models for hardware-efficient inference.

This paper introduces an instruction-preserving decoding strategy to prevent visual contrastive decoding from discarding valid abstention responses.

Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits.

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

The paper introduces a domain-general framework using predictive world models to guide action selection and policy improvement in recursive self-improvement settings.

The paper introduces a benchmark assessing language models on operational planning from incomplete, complex, and contradictory military document sets.

Proposes a text anchor re-grounding method using low-rank adapters to improve concept recognition in Segment Anything Model 3.

This paper proposes a low-rank bypass to accumulate small updates and reduce compounding errors in quantized recurrent model states.

Replaces free-form reasoning text with executable computational graphs trained via reinforcement learning to improve reasoning speed and verifiability.

Introduces SwitchMem, a routing system that dynamically selects between text and visual memory representations for vision-language GUI agents.

The paper introduces a multi-agent system to automate data curation and dataset editing for supervised fine-tuning.

This paper presents a technology-neutral reference architecture for integrating agentic AI capabilities into enterprise software systems.

The paper investigates progressive depth growth during self-supervised vision transformer pretraining to improve downstream accuracy while reducing computation.

Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images.

The authors formulate a meta-game framework to dynamically route and select language models during strategic multi-agent interactions.

This paper proposes an agentic video framework combining change-aware memory and targeted replay for streaming video understanding.

This paper formulates a trace-minimization framework to design causal error-feedback coefficients for neural network weight quantization.

The paper applies continuous bitstream diffusion to joint and conditional generation across multimodal data without learned tokenizers.

The paper presents an autonomous agent architecture that distills distributed system incident resolutions into reusable detectors and operational playbooks.

Presents LeanPRM, a method using the Lean proof assistant to generate step-level correctness labels for training natural-language mathematical proof verifiers.

The paper introduces a Persian vision-language-action dataset and benchmark to evaluate robotic instruction grounding under deliberate ambiguity and infeasibility.

This paper demonstrates a security risk where dormant backdoors trigger safety failures only after merging models that passed individual safety audits.

Introduces DeltaTTG, a method using predicted remaining token count as a progress signal to allocate advantages in multi-turn reinforcement learning.

The paper introduces a benchmark and model for evaluating program correctness and generating diagnostic counterexamples for competitive programming code.

Proposes an attention mechanism that switches from linear complexity during prefilling to full attention during token decoding.

This paper evaluates multilingual bias audits across ten language models, showing that lower stereotype scores often reflect translation artifacts rather than fairer behavior.

This paper introduces a training framework that enables harness-based software agents to self-improve without relying on external teacher models.

Presents an environment scaling framework that alters dependencies to scale agent benchmark difficulty while preserving task objectives.

Proposes BoundaryEvo, dynamically adapting training data around a student model's evolving capability boundary during long-chain reasoning distillation.

Pre-trains embodied robot policies on human egocentric videos using learned continuous latent action representations.

Connects penetration-testing language model agents to graph databases to enable multi-step exploit-chain reasoning.

The paper introduces an inference-time guidance method that enforces temporal logic constraints during text-to-video generation.

Proposes a severity-weighted character error rate that charges text transcription errors based on their practical impact on reader understanding.

The paper studies how calibration dataset size, language, and data distribution affect post-training expert pruning in mixture-of-experts models.

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

The paper introduces a benchmark to evaluate whether language model agents reliably delete sensitive user data and secret evidence from memory upon request.

The paper proposes an agent architecture that dynamically switches between latent reasoning and text interaction while maintaining external tool access.

Proposes category-conditioned expert routing biases to mitigate data imbalance in mixture-of-experts language models.

Transfers editing capabilities from image pairs to video diffusion models using in-context visual demonstrations across decomposed sub-tasks.

Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire.

Presents an evaluation framework combining simulated mental health patients and expert-developed rubrics to assess conversational language models.

The paper introduces a multi-domain benchmark to evaluate language model agents on identifying, understanding, and applying domain-specific rules across complex files.

The paper describes an automated system using language models to verify corporate loan applications against internal bank lending policies.

Introduces a multilingual tokenizer that uses structural discovery and linguistic constraints to improve downstream model representations.

Proposes an adaptive multi-source steering decoding method to mitigate hallucinations in vision-language models by balancing visual and textual context contributions.

The authors present an engine to synthesize environments, tasks, and rubrics for training and evaluating data analysis agents.

The paper proposes an interface framework that tracks and controls application state changes made by graphical user interface agents.

The paper introduces benchmarks and summarizes winning solutions for city-scale audio-visual reasoning and spatial intelligence at the Perception Test 2026.

Presents a hypothesis testing framework using shadow models to audit whether a language model was distilled from a proprietary teacher.

Introduces a training-free inference method using Gaussian optimal transport to reduce object hallucination in vision-language models without suppressing valid representations.