2/25/2026
Show HN: AgentPass – Identity layer for AI agents (passports, email, trust)
AgentPass provides cryptographic identity and authentication infrastructure for autonomous AI agents. Identity layer enabling agent autonomy.
2/25/2026
Agent context management: ephemeral vs. durable classification
Relay product for efficient agent context management via ephemeral/durable classification. Token optimization for autonomous agents.
2/25/2026
AI_ATTRIBUTION.md: A Standard for Tracking Creative Control in Human-AI Coding
Standard for tracking human-AI creative control attribution in code generation. Developer tool for documentation and collaboration.
2/25/2026
vLLM WideEP and Large-Scale Serving Toward Maturity on Blackwell (Part I)
vLLM optimization achieving 26.2K prefill and 10.1K decode throughput on NVIDIA Blackwell for DeepSeek MoE models. Production LLM serving research.
2/25/2026
Webgrid Eval: LLM vision + tool-use on Neuralink's cursor control task
Benchmark for evaluating LLM vision and tool-use on cursor control task based on Neuralink's Webgrid test. Measures multimodal agent capabilities.
2/25/2026
I rebuilt Game Boy on web using 1 prompt and 5 parallel agents in 48 hours
Multi-agent autonomous workflow built Game Boy GBA emulator in 48 hours using TypeScript. Demonstrates agent orchestration and tool use.
2/25/2026
DataClaw: Publish your Claude Code chats to HuggingFace with a single command
DataClaw tool exports Claude Code conversation history to HuggingFace datasets. Data extraction tool with open source intent.
2/25/2026
Anthropic accuses DeepSeek and other Chinese rivals of mass data theft
News article about Anthropic accusing Chinese AI firms of model distillation for IP theft. Industry news, not technical content.
2/25/2026
An artificial intelligence framework for end-to-end rare disease phenotyping from clinical notes using large language models
LLM-based system for extracting clinical phenotypes from medical notes and standardizing to Human Phenotype Ontology terms for rare disease diagnosis.
2/25/2026
DMCD: Semantic-Statistical Framework for Causal Discovery
Framework combining LLM-based semantic analysis of metadata with statistical causal discovery via conditional independence testing.
2/25/2026
Diffusion Modulation via Environment Mechanism Modeling for Planning
Diffusion models for trajectory generation in offline reinforcement learning, improving consistency between generated transitions and environment dynamics.
2/25/2026
Implicit Intelligence -- Evaluating Agents on What Users Don't Say
Benchmark for evaluating AI agents on implicit requirements (accessibility, privacy, risks) beyond explicit instructions, addressing real-world underspecified requests.
2/25/2026
Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use
Research on improving tool descriptions for LLM agents. Addresses tool interface optimization as bottleneck for reliable agent tool use.
2/25/2026
PreScience: A Benchmark for Forecasting Scientific Contributions
PreScience benchmark for forecasting scientific advances using AI. Tests if models trained on historical research can predict future collaborators, impactful directions, and emerging problems.
2/25/2026
KairosVL: Orchestrating Time Series and Semantics for Unified Reasoning
KairosVL integrates time series analysis with semantic reasoning using reinforcement learning for complex temporal prediction tasks combining numerical and contextual understanding.
2/25/2026
ActionEngine: From Reactive to Programmatic GUI Agents via State Machine Memory
ActionEngine framework enables GUI agents to operate programmatically via state machine memory instead of reactive step-by-step vision-language model calls, reducing costs and improving accuracy.
2/25/2026
Inner Speech as Behavior Guides: Steerable Imitation of Diverse Behaviors for Human-AI coordination
Research on imitation learning for AI agents to exhibit diverse human-like behaviors and adapt to contexts. Uses inner speech mechanisms for steering agent behavior in human-AI coordination tasks.
2/25/2026
From Logs to Language: Learning Optimal Verbalization for LLM-Based Recommendation in Production
Data-centric framework that learns optimal verbalization of user interaction logs for LLM-based recommendation systems, replacing rigid template concatenation.
2/25/2026
CausalReasoningBenchmark: A Real-World Benchmark for Disentangled Evaluation of Causal Identification and Estimation
CausalReasoningBenchmark evaluates causal inference systems separately on identification and estimation steps across 173 real-world queries and 138 datasets.
2/25/2026
Physics-based phenomenological characterization of cross-modal bias in multimodal models
Characterizes cross-modal bias in multimodal LLMs from algorithmic fairness perspective, examining comparative and non-comparative fairness contexts.
2/25/2026
When can we trust untrusted monitoring? A safety case sketch across collusion strategies
Examines safety guarantees for untrusted monitoring deployments where misaligned AI models oversee each other across different collusion strategies.
2/25/2026
Identifying two piecewise linear additive value functions from anonymous preference information
Proposes elicitation procedure for identifying two piecewise linear additive value functions from anonymous preference pairs without knowing decision-maker correspondence.
2/25/2026
Grounding LLMs in Scientific Discovery via Embodied Actions
EmbodiedAct framework grounds LLM scientific reasoning in physical simulation with runtime perception to detect anomalies and enable interactive discovery.
2/25/2026
Recursive Belief Vision Language Model
Recursive Belief Vision Language Model (RBVLM) maintains hidden state for long-horizon manipulation under partial observability, reducing action repetition and inference latency.
2/25/2026
How Foundational Skills Influence VLM-based Embodied Agents:A Native Perspective
Evaluates foundational skills required for VLM-based embodied agents in native control settings with both low and high-level task benchmarks.
2/25/2026
PromptCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding
PromptCD applies contrastive decoding at test time to improve LLM alignment with human preferences without additional training data or computational costs.
2/25/2026
Online Algorithms with Unreliable Guidance
Introduces online algorithms with unreliable guidance (OAG) model for ML-augmented online decision making that separates predictive and algorithmic components.
2/25/2026
ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction
ICON is an inference-time defense framework for LLM agents against indirect prompt injection attacks that mitigates threats while preserving valid agentic workflows.
2/25/2026
Counterfactual Simulation Training for Chain-of-Thought Faithfulness
Counterfactual Simulation Training (CST) improves chain-of-thought faithfulness by rewarding reasoning paths that accurately predict model outputs versus counterfactual scenarios.
2/25/2026
Buffer Matters: Unleashing the Power of Off-Policy Reinforcement Learning in Large Language Model Reasoning
Introduces BAPO, an off-policy reinforcement learning framework with verifiable rewards that improves data efficiency in LLM post-training through adaptive buffer management.
2/25/2026
Modality-Guided Mixture of Graph Experts with Entropy-Triggered Routing for Multimodal Recommendation
Proposes mixture of graph experts with entropy-triggered routing for multimodal recommendation systems to address modality imbalance in sparse feedback settings.
2/25/2026
Balancing Multiple Objectives in Urban Traffic Control with Reinforcement Learning from AI Feedback
Uses RL from AI feedback (RLAIF) with LLMs to generate preference labels for multi-objective urban traffic control, reducing reliance on human annotation.
2/25/2026
CHESS: Context-aware Hierarchical Efficient Semantic Selection for Long-Context LLM Inference
CHESS proposes context-aware token pruning for long-context LLM inference that reduces KV cache overhead while maintaining quality through step-wise relevance and semantic awareness.
2/25/2026
PyVision-RL: Forging Open Agentic Vision Models via RL
PyVision-RL is an open-source reinforcement learning framework for training multimodal agentic models that prevents interaction collapse and sustains multi-turn reasoning via tool rewards.
2/25/2026
Pipeline for Verifying LLM-Generated Mathematical Solutions
Pipeline for automatic and interactive verification of LLM-generated mathematical solutions as alternative to answer-only evaluation, enabling formal and informal solution generation.
2/25/2026
POMDPPlanners: Open-Source Package for POMDP Planning
POMDPPlanners is an open-source Python package for evaluating POMDP planning algorithms with benchmarks, hyperparameter optimization, and parallel simulation capabilities.
2/25/2026
Qwen-BIM: developing large language model for BIM-based design with domain-specific benchmark and dataset
Qwen-BIM proposes an LLM fine-tuned for building information modeling with domain-specific benchmark and dataset to improve performance on BIM-based design tasks.
2/25/2026
Pressure Reveals Character: Behavioural Alignment Evaluation at Depth
Alignment benchmark with 904 realistic multi-turn scenarios across six categories evaluating LLM behavior under pressure including honesty, safety, and scheming.
2/25/2026
Diagnosing Causal Reasoning in Vision-Language Models via Structured Relevance Graphs
Vision-Language Causal Graphs enable structured evaluation of causal reasoning in LVLMs, distinguishing between reasoning failures and spurious correlations.
2/25/2026
Predicting Sentence Acceptability Judgments in Multimodal Contexts
Study examines how visual context affects human and LLM predictions of sentence acceptability compared to textual context.
2/25/2026
HELP: HyperNode Expansion and Logical Path-Guided Evidence Localization for Accurate and Efficient GraphRAG
HELP improves GraphRAG efficiency and accuracy by expanding hypernodes and using logical path-guided evidence localization for multi-hop reasoning.
2/25/2026
Architecting AgentOS: From Token-Level Context to Emergent System-Level Intelligence
AgentOS is a conceptual framework redefining LLMs as dynamic autonomous cognitive systems bridging token-level processing to system-level intelligence.
2/25/2026
LogicGraph : Benchmarking Multi-Path Logical Reasoning via Neuro-Symbolic Generation and Verification
LogicGraph benchmark evaluates LLMs on multi-path logical reasoning with neuro-symbolic generation and verification beyond single correct proofs.
2/25/2026
Tool Building as a Path to "Superintelligence"
Benchmark measuring step-success probability for LLM superintelligence via test-time search using GF(2) circuit reconstruction tasks.
2/25/2026
Motivation is Something You Need
Novel training paradigm inspired by affective neuroscience using motivation states to activate a larger model intermittently alongside continuous base model training.
2/25/2026
The Initial Exploration Problem in Knowledge Graph Exploration
Paper addresses the initial exploration problem for lay users encountering unfamiliar Knowledge Graphs without semantic web expertise.
2/25/2026
A Benchmark for Deep Information Synthesis
DEEPSYNTH benchmark evaluates LLM-based agents on complex tasks requiring multi-source information synthesis and tool use like web browsing and code execution.
2/25/2026
CG-DMER: Hybrid Contrastive-Generative Framework for Disentangled Multimodal ECG Representation Learning
CG-DMER combines contrastive and generative learning for multimodal ECG representation learning with clinical reports.
2/25/2026
NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning
NoRD is a data-efficient Vision-Language-Action model for autonomous driving that achieves competitive performance without dense reasoning annotations.
2/25/2026
Aletheia tackles FirstProof autonomously
Aletheia, an AI agent powered by Gemini 3 Deep Think, autonomously solved 6 out of 10 FirstProof mathematical problems.