4/6/2026
Moondream Segmentation: From Words to Masks
Vision-language model extension for referring image segmentation using autoregressive decoding and reinforcement learning refinement.
4/6/2026
Making Written Theorems Explorable by Grounding Them in Formal Representations
System grounding LLM-generated explanations in formal representations to enable interactive exploration of mathematical proofs.
4/6/2026
LitPivot: Developing Well-Situated Research Ideas Through Dynamic Contextualization and Critique within the Literature Landscape
Tool for developing research ideas through dynamic literature contextualization and critique using LLMs.
4/6/2026
Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents
Security analysis of memory-based LLM web agents, demonstrating environment-injected poisoning attacks through persistent memory exploitation.
4/6/2026
Analytic Drift Resister for Non-Exemplar Continual Graph Learning
Continual graph learning method addressing feature drift in non-exemplar settings using analytic continual learning.
4/6/2026
GBQA: A Game Benchmark for Evaluating LLMs as Quality Assurance Engineers
Game benchmark with 124 bugs for evaluating LLMs' ability to autonomously discover bugs as QA engineers in dynamic environments.
4/6/2026
Communication-free Sampling and 4D Hybrid Parallelism for Scalable Mini-batch GNN Training
Distributed training approach for graph neural networks using communication-free sampling and hybrid parallelism.
4/6/2026
Generalization Limits of Reinforcement Learning Alignment
Theoretical analysis of reinforcement learning alignment limitations in LLMs, demonstrating generalization failures through compound jailbreak attacks.
4/6/2026
Low-Rank Compression of Pretrained Models via Randomized Subspace Iteration
Efficient model compression using randomized subspace iteration for low-rank decomposition of pretrained models.
4/6/2026
Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems
Study of sycophancy propagation in multi-agent LLM systems, examining how agents' awareness of others' biases affects collaborative discussions.
4/6/2026
Do Agent Societies Develop Intellectual Elites? The Hidden Power Laws of Collective Cognition in LLM Multi-Agent Systems
Large-scale empirical study of coordination dynamics in LLM multi-agent systems, analyzing scaling behavior and power laws in collective cognition.
4/6/2026
Eligibility-Aware Evidence Synthesis: An Agentic Framework for Clinical Trial Meta-Analysis
Agentic framework using LLMs for automated clinical trial evidence synthesis and meta-analysis with eligibility-aware study selection.
4/6/2026
Finding Belief Geometries with Sparse Autoencoders
Using sparse autoencoders to understand geometric structure of belief representations in transformer models and LLMs.
4/6/2026
Beyond Semantic Manipulation: Token-Space Attacks on Reward Models
Token-space adversarial attacks on reward models used in RLHF, introducing token mapping perturbation attack paradigm beyond semantic manipulation.
4/6/2026
Efficient3D: A Unified Framework for Adaptive and Debiased Token Reduction in 3D MLLMs
Framework for reducing computational overhead in 3D multimodal LLMs through adaptive token reduction for resource-constrained deployment.
4/6/2026
DocShield: Towards AI Document Safety via Evidence-Grounded Agentic Reasoning
AI agent system for document forgery detection using evidence-grounded reasoning, combining detection, localization, and explanation for document safety.
4/6/2026
Trivial Vocabulary Bans Improve LLM Reasoning More Than Deep Linguistic Constraints
Controlled replication study examining vocabulary constraints versus linguistic structures in LLM reasoning, testing E-Prime effects on cognition.
4/6/2026
Evaluating the Formal Reasoning Capabilities of Large Language Models through Chomsky Hierarchy
Systematic evaluation framework for LLM formal reasoning capabilities using Chomsky hierarchy and computational complexity theory.
4/6/2026
V2X-QA: A Comprehensive Reasoning Dataset and Benchmark for Multimodal Large Language Models in Autonomous Driving Across Ego, Infrastructure, and Cooperative Views
Multimodal LLM benchmark for autonomous driving with vehicle, infrastructure, and cooperative viewpoints, evaluating reasoning across V2X conditions.
4/6/2026
IndustryCode: A Benchmark for Industry Code Generation
Multi-domain benchmark for industry code generation across finance, automation, and aerospace using LLMs, addressing single-domain limitations.
4/6/2026
Random Is Hard to Beat: Active Selection in online DPO with Modern LLMs
Evaluation of active preference learning versus random sampling in online DPO for modern LLMs, showing random sampling is surprisingly competitive.
4/6/2026
SentinelAgent: Intent-Verified Delegation Chains for Securing Federal Multi-Agent AI Systems
Formal framework for verifiable delegation chains in multi-agent AI systems, defining properties for authorization tracking and policy enforcement.
4/6/2026
Disrupting Cognitive Passivity: Rethinking AI-Assisted Data Literacy through Cognitive Alignment
Framework for improving data literacy in AI-assisted analysis by disrupting cognitive passivity through guided reasoning rather than direct answers.
4/6/2026
Rubrics to Tokens: Bridging Response-level Rubrics and Token-level Rewards in Instruction Following Tasks
Rubric-based RL framework bridging response-level and token-level rewards for LLM alignment in instruction following tasks.
4/6/2026
ChatSVA: Bridging SVA Generation for Hardware Verification via Task-Specific LLMs
Task-specific LLM framework for generating SystemVerilog assertions for hardware verification, addressing data scarcity and accuracy challenges.
4/6/2026
QAPruner: Quantization-Aware Vision Token Pruning for Multimodal Large Language Models
Quantization-aware vision token pruning for multimodal LLMs, optimizing coupled compression techniques for resource-constrained deployment.
4/6/2026
Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis
First comprehensive security analysis of Agent Skills, an open standard for modular LLM agent packages, covering threat taxonomy and vulnerabilities.
4/6/2026
LLM+Graph@VLDB'2025 Workshop Summary
Workshop on integrating LLMs with graph-structured data, covering algorithms and systems for bridging LLMs, graph databases, and ML for practical applications.
4/6/2026
One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging
Study of weight-space model merging for multilingual machine translation, evaluating behavior when combining independently fine-tuned models.
4/6/2026
Toward an Artificial General Teacher: Procedural Geometry Data Generation and Visual Grounding with Vision-Language Models
Procedural geometry data generation and visual grounding using vision-language models for geometry education as referring image segmentation.
4/6/2026
Corporations Constitute Intelligence
Legal analysis of Anthropic's AI constitution document as governance framework, discussing limitations in military and surveillance contexts.
4/6/2026
Council Mode: Mitigating Hallucination and Bias in LLMs via Multi-Agent Consensus
Council Mode: multi-agent consensus approach mitigating hallucinations and bias in MoE LLMs through coordinated expert activation.
4/6/2026
Learning from Synthetic Data via Provenance-Based Input Gradient Guidance
Learning method using provenance-based input gradient guidance to improve model discrimination robustness with synthetic training data.
4/6/2026
LogicPoison: Logical Attacks on Graph Retrieval-Augmented Generation
LogicPoison attacks exploiting logical vulnerabilities in Graph-RAG systems that ground LLM reasoning in knowledge graphs.
4/6/2026
Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference
Measuring latency and quality tradeoffs of prompt compression techniques for accelerating LLM inference in RAG systems.
4/6/2026
Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
Mitigating reward hacking in RLHF by analyzing and correcting flipped advantage signs in reward model parameters.
4/6/2026
Self-Optimizing Multi-Agent Systems for Deep Research
Self-optimizing multi-agent system for deep research that iteratively plans, retrieves, and synthesizes evidence across documents.
4/6/2026
FedSQ: Optimized Weight Averaging via Fixed Gating
FedSQ algorithm optimizing weight averaging in federated learning across heterogeneous client data with fixed gating mechanisms.
4/6/2026
R2-Write: Reflection and Revision for Open-Ended Writing with Deep Reasoning
R2-Write framework exploring deep reasoning with chain-of-thought for open-ended writing tasks using reasoning models.
4/6/2026
Beyond Isolated Tasks: A Framework for Evaluating Coding Agents on Sequential Software Evolution
SWE-STEPS dataset and framework for evaluating coding agents on sequential, long-horizon software development tasks with accumulated technical debt.
4/6/2026
JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency
JoyAI-LLM Flash, an efficient mixture-of-experts mid-scale LLM with 20 trillion token pretraining optimized for token efficiency.
4/6/2026
Querying Structured Data Through Natural Language Using Language Models
Open-source methodology enabling natural language queries on structured data by training LLMs to generate executable queries with synthetic training data.
4/6/2026
Verbalizing LLMs' assumptions to explain and control sycophancy
Framework for eliciting and verbalizing LLM assumptions to explain and mitigate sycophancy behavior in user interactions.
4/6/2026
Credential Leakage in LLM Agent Skills: A Large-Scale Empirical Study
Large-scale empirical study of credential leakage vulnerabilities in 17,022 LLM agent skills, identifying 520 vulnerable skills with taxonomy of 10 leakage patterns.
4/6/2026
Supply-Chain Poisoning Attacks Against LLM Coding Agent Skill Ecosystems
Security study of supply-chain poisoning attacks against LLM coding agents through malicious third-party skills with system-level execution.
4/6/2026
Co-Evolution of Policy and Internal Reward for Language Agents
Self-Guide method for co-evolving policy and internal reward in LLM agents, addressing sparse reward bottleneck in long-horizon training.
4/6/2026
AlertStar: Path-Aware Alert Prediction on Hyper-Relational Knowledge Graphs
Knowledge graph completion approach for network alert prediction modeling cyber-attacks as hyper-relational statements.
4/6/2026
Can VLMs Truly Forget? Benchmarking Training-Free Visual Concept Unlearning
Benchmarking training-free unlearning methods for removing sensitive visual concepts from vision-language models.
4/6/2026
An Independent Safety Evaluation of Kimi K2.5
Safety evaluation of Kimi K2.5 open-weight LLM assessing CBRNE misuse, cybersecurity, alignment, and bias risks.
4/6/2026
Domain-Adapted Retrieval for In-Context Annotation of Pedagogical Dialogue Acts
Domain-adapted RAG pipeline using fine-tuned embedding models for pedagogical dialogue act annotation without generative model fine-tuning.