All papers

Proposes a reinforcement learning credit-assignment method for document parsing models to align rewards with structured page elements.

The authors introduce a benchmark to evaluate autonomous scientific discovery agents making iterative experimental decisions under fixed budgets.

This paper analyzes source-side training dynamics to prospectively detect out-of-distribution performance degradation before evaluation.

Develops a retrieval-enhanced routing framework that dispatches multimodal queries to specialized vision-language models based on local failure patterns.

Introduces a step-level memory reuse framework using transition graphs to adapt relevant prior mobile graphical interface execution trajectories.

This paper introduces a benchmark and rule-based metric for generating and evaluating audio descriptions in university slide lectures.

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

Evaluates whether temporal illumination inconsistency priors improve face-swap deepfake detection performance across diverse video datasets.

Linearizes multilayer perceptrons using low-degree rational approximations of activation functions to reduce inference costs across 3D models and language models.

Presents a framework that defers memory construction to test time to answer queries across long interaction histories.

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

The paper synthesizes failure modes, evaluation methods, and defense protocols across the lifecycle of world models.

Proposes a framework that encodes visual trajectories into reusable skill cards to reinforce vision-language agent planning and rewards.

Introduces an end-to-end open pipeline, pretraining corpus, and family of decoder-only models trained from scratch for Romanian.

Presents ME-World, an egocentric world model generating synchronized multi-agent first-person video streams with fine-grained physical interactions.

This paper presents an evaluation benchmark and guardrail framework for detecting safety risks in multi-turn interactions with AI companions.

Proposes a dynamic structural pruning method using biologically inspired plasticity to reduce neural network parameters during training.

The paper presents a framework combining retrieval-augmented generation and structured prompts to generate context-aware malware deception playbooks.

Develops a framework for egocentric multimodal assistants to decide both when to speak and what guidance to provide.

Introduces Compo and a spatial canvas interface enabling users or agents to guide poster layout and image generation.

This paper proposes decoupling error proposing from error validation to filter invalid candidates under budget constraints for expert human review.

Introduces a dynamic monitoring framework that adaptively schedules runtime inspections and intervention guidance for coding agents.

Introduces an attention-aligned recurrent architecture that maintains persistent state representations across sequences using rolling low-dimensional traces.

Introduces a preview-guided admission framework allowing search agents to selectively incorporate retrieved candidates into the ongoing reasoning context.

Introduces a dataset and benchmark to model multimodal emotion understanding through cognitive appraisal reasoning rather than superficial affective cues.

The paper formalizes layer-to-layer representation sharing as a memory storage and retrieval framework for deep transformer networks.

This paper introduces a benchmark of expert-annotated question-answer pairs to evaluate multimodal comprehension of long historical Chinese documents.

This paper introduces an evaluation framework and reward models to assess how well search agents ground generated answers in retrieved, independent sources.

The paper proposes a framework that uses reasoning graphs to locate claim-level disagreements across sampled traces and correct multi-step language model reasoning.

Presents a temporal memory architecture using hierarchical tree structures to trace evolving multi-party conversational histories for agents.

This research statement outlines a conceptual framework and future directions for robot scene understanding under incomplete sensory and environmental data.

The paper proposes a learnable low-rank polynomial sketch to reduce approximation errors in linear attention mechanisms.

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

The paper audits historical evaluation tasks to see whether an agent's internal post-cutoff memory skews its point-in-time evidence gathering.

Generates synthetic multi-agent dialogues to benchmark conversational agent robustness against non-cooperative and adversarial user behaviors.

The paper introduces a speculative decoding framework that uses early recurrent states to draft tokens in looped transformers.

This paper proposes a progressive layout generation framework that dynamically adapts graphic design compositions to the rendered canvas.

The paper introduces a benchmark evaluating how multi-agent teams develop communication protocols under asymmetric information and restricted communication channels.

The paper demonstrates that generative recommenders are vulnerable to targeted poisoning attacks using fewer than 1% malicious synthetic interactions.

This paper presents a benchmark to evaluate performance, resource usage, and privacy trade-offs when language model agents collaborate across edge and cloud environments.

The paper discusses visual budgeting strategies to manage token redundancy for scalable multi-image reasoning in multimodal language models.

The paper presents a framework to probabilistically localize user-intent misinterpretations and repair conversational states in multi-turn assistant systems.

This paper fine-tunes a quantized decision model to classify diseases from patient symptom descriptions without free-form text generation.

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

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

The paper introduces a visual-memory framework enabling unified multimodal models to handle long-horizon interleaved text and image generation.

The paper investigates how observing partner states, sharing knowledge, and predicting partner actions affect coordination in embodied multi-agent teams.

Introduces Socratic Skill Learning, interleaving expert guidance with student practice to produce reusable and transferable agent skills across tasks.

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

Proposes stabilization modules to enable half-precision FP16 proximal policy optimization training on commodity GPUs without numerical divergence.

The paper uses EEG data to compare how image-text aligned and vision-only models align with developing human visual representations.

Applies regenerative software primitives so programs can be deleted and rebuilt by agents without losing architectural intent and knowledge.

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

Formulates machine unlearning for LLM agents across parametric memory, external persistent storage, and tool-accessible environments.

Replaces dense LoRA experts with block-diagonal factorizations to isolate gradients and prevent catastrophic forgetting during continual multimodal tuning.

The paper presents a framework aligning visual image restoration with navigation decisions for vision-language navigation in degraded indoor fire environments.

The paper presents SparPR, a runtime key-value cache recovery mechanism that reduces memory bottlenecks during model inference on consumer hardware.

The paper introduces an optimization method that separates error identification from credit assignment to improve reinforcement learning for tool-using agents.

The paper presents a benchmark to evaluate whether formal specifications generated by language models accurately reflect natural-language requirements.

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