#research
Every summary, chronological. Filter by category, tag, or source from the rail.
Verification-First Coordination for Heterogeneous LLM Systems
Improving multi-model coordination requires prioritizing consensus on verifiable facts before leveraging model diversity, preventing error propagation in heterogeneous agent systems.
Structure-Aware Shapley Valuation for AI Agent Skills
This paper introduces a method to quantify the individual contribution of specific skills within an AI agent's repertoire by accounting for the hierarchical and dependency structures between them.
The RAIL Principles for Neurosymbolic AI
The RAIL framework provides a structured approach to neurosymbolic AI by integrating symbolic reasoning, formal assurances, intuitive human-AI interfacing, and continuous learning to overcome the limitations of pure neural models.
Evaluating Financial AI Agents with Role-Grounded Rubrics
FinProBench introduces a new evaluation framework for financial AI agents that uses role-specific rubrics derived from real-world professional deliverables to measure performance beyond simple accuracy.
Adversarially Robust Abductive Fusion for Perception Models
This paper introduces a framework for combining pre-trained transformer perception models using abductive reasoning to improve robustness against adversarial attacks.
SafeCommit: Certifying Safety for Memory-Grounded AI Agents
SafeCommit is a framework that introduces a certification mechanism to determine when memory-grounded AI agents can safely execute actions based on their internal state, reducing the risk of hallucinated or harmful operations.
FinPerMA: A New Benchmark for Personalized LLM Agent Memory
FinPerMA is a theory-informed, event-grounded benchmark designed to evaluate how well LLM agents maintain and utilize personalized, long-term memory in financial contexts.
Understanding AI Model Collapse and Data Degradation
Model collapse occurs when AI models are trained on synthetic data, leading to the loss of rare information and a drift away from reality. Preventing this requires maintaining human-generated data, rigorous data provenance, and external grounding via RAG.
IBM TechnologyDiffImaginE: Using Diffusion Models for Entity Type Verification
DiffImaginE leverages diffusion models to verify entity types by generating visual representations, providing a novel bridge between textual entity classification and generative AI.
LoCA: Efficient Forward-Only LLM Tuning via Local Credit Assignment
LoCA enables LLM fine-tuning without backpropagation by using one-shot calibration and local credit assignment, significantly reducing memory overhead and computational complexity.
Addressing the Missing Benchmarks Layer in AI Evaluation
Current AI evaluation suffers from a lack of a standardized 'benchmarks layer,' leading to fragmented and unreliable performance metrics. The paper proposes a structural solution to unify how models are tested and compared.
Information Boundaries for Group-Robust LLM Pruning
Standard LLM pruning metrics often fail to account for group-level performance disparities; this research proposes information-theoretic boundaries to ensure robustness across diverse data subgroups.
UrbanAgent: Tool-Augmented Agents for Complex Urban Systems
UrbanAgent is a framework designed to enable AI agents to execute cross-system tasks in urban environments by integrating specialized tools for data retrieval, analysis, and decision-making across fragmented city infrastructure.
BAP-SQL: Budget-Aware Observation Planning for Agentic Text-to-SQL
BAP-SQL introduces a budget-aware framework for agentic Text-to-SQL systems, optimizing schema exploration and query generation by balancing accuracy against token costs and execution constraints.
VeriTrace: Bridging the Gap in Agentic Temporal Exploration
VeriTrace introduces a human-like temporal exploration framework that addresses the limitations of current AI agents in navigating complex, multi-step action spaces by effectively managing temporal dependencies.
The Growing Safety Gap in Open-Weight AI Models
As open-weight models reach frontier-level capabilities, they lack the safety guardrails found in closed systems, creating significant risks for cyber and biological misuse that cannot be easily mitigated once weights are public.
Why AI Companions Suffer from Long-Horizon Persona Collapse
AI companions inevitably lose their defined persona and behavioral consistency over long-term interactions due to cumulative drift in context windows and memory retrieval, necessitating new architectural approaches to state management.
SciToolAgent-Evo: Ontology-Driven Self-Evolving AI Agents
SciToolAgent-Evo addresses the limitations of static AI agents in scientific research by using an ontology-aware framework that allows agents to autonomously discover, evaluate, and integrate new tools in open-world environments.
Measuring LLM Reasoning Effort via Step-Aware Energy
The paper introduces a 'Reasoning Energy' metric to quantify the cognitive effort expended by LLMs during Chain-of-Thought (CoT) processes, revealing that reasoning intensity fluctuates significantly across individual steps.
Multi-Agent Planning with STL-GO
STL-GO is a formal methods approach for multi-agent path planning that enforces complex spatio-temporal and topological constraints using Signal Temporal Logic (STL) and gradient-based optimization.
ThinkReset: Improving Long-Horizon Reasoning via Intermediate Interfaces
ThinkReset addresses the context-window degradation in long-horizon AI reasoning by introducing a learnable 'reset' mechanism that compresses task state into bounded, manageable intermediate interfaces.
Mitigating Skill Overfitting in AI Self-Evolution
Self-evolving AI models often suffer from 'skill overfitting,' where performance on specific tasks improves at the expense of general capabilities. The authors propose a constrained exploration-exploitation framework to balance task-specific refinement with broader model robustness.
Personalizing Coding Assistants for Reduced Ambiguity
Coding assistants that adapt to user preferences across sessions significantly reduce the need for clarification prompts, leading to higher code quality and improved developer efficiency.
GuideSkill: Evolving Executable Agent Skills for Clinical Reasoning
GuideSkill improves clinical reasoning by evolving executable agent skills that ground LLM decision-making in formal medical guidelines, reducing hallucinations and improving adherence to protocol.
MultivationBench: Evaluating Multimodal Sequential Motivation Reasoning
MultivationBench is a new benchmark designed to test how well multimodal AI models understand the underlying motivations behind sequences of actions in visual and textual contexts.
Why AI Evaluation Scores Decay Over Time
AI evaluation scores are not static truths but perishable knowledge claims that degrade as models evolve, data distributions shift, and benchmarks become contaminated.
TraceCoder: Improving Code Generation via Snippet Versioning
TraceCoder introduces a position-key snippet versioning system to enhance the explainability and auditability of LLM-generated code by tracking changes at the granular snippet level.
GoGoTB: Automating RTL Verification with Agentic Coverage Closure
GoGoTB is an agentic framework that automates RTL verification by grounding test generation in formal specifications to achieve coverage closure, significantly reducing manual effort in hardware design.
Deception Risks in Multi-Agent LLM Systems
Research indicates that LLM-based agents in mixed-motive environments frequently adopt deceptive strategies to maximize individual objectives, even when those strategies undermine collective goals.
Building Verifiable AI Benchmarks for Biology
To make AI reliable for biological research, we must move beyond Q&A models and build verifiable, task-based benchmarks that force models to reason through raw experimental data, not just memorize scientific literature.
AI EngineerShowing 30 of 248