Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Rich Vehicle Routing Problem in Disaster Management enabling Temporally-causal Transhipments across Multi-Modal Transportation Network

Omni-CLST: Error-aware Curriculum Learning with guided Selective chain-of-Thought for audio question answering

Humor in Pixels: Benchmarking Large Multimodal Models Understanding of Online Comics

Enhancing Generalization in Vision-Language-Action Models by Preserving Pretrained Representations

The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models

Standards in the Preparation of Biomedical Research Metadata: A Bridge2AI Perspective

Visible Yet Unreadable: A Systematic Blind Spot of Vision Language Models Across Writing Systems

Zero-Knowledge Proofs in Sublinear Space

Emergent Social Dynamics of LLM Agents in the El Farol Bar Problem

Training Text-to-Molecule Models with Context-Aware Tokenization

Towards Trustworthy Vital Sign Forecasting: Leveraging Uncertainty for Prediction Intervals

Assessing Large Language Models on Islamic Legal Reasoning: Evidence from Inheritance Law Evaluation

Language Models Identify Ambiguities and Exploit Loopholes

Mining the Long Tail: A Comparative Study of Data-Centric Criticality Metrics for Robust Offline Reinforcement Learning in Autonomous Motion Planning

GWM: Towards Scalable Gaussian World Models for Robotic Manipulation

ECHO: Frequency-aware Hierarchical Encoding for Variable-length Signals

Hierarchical Evaluation Function: A Multi-Metric Approach for Optimizing Demand Forecasting Models

Posterior-GRPO: Rewarding Reasoning Processes in Code Generation

Legal Knowledge Graph Foundations, Part I: URI-Addressable Abstract Works (LRMoo F1 to schema.org)

Pareto-Grid-Guided Large Language Models for Fast and High-Quality Heuristics Design in Multi-Objective Combinatorial Optimization

Self-supervised learning on gene expression data

EdgeProfiler: A Fast Profiling Framework for Lightweight LLMs on Edge Using Analytical Model

Sarc7: Evaluating Sarcasm Detection and Generation with Seven Types and Emotion-Informed Techniques

MythTriage: Scalable Detection of Opioid Use Disorder Myths on a Video-Sharing Platform

Calibrating LLMs for Text-to-SQL Parsing by Leveraging Sub-clause Frequencies

Puzzled by Puzzles: When Vision-Language Models Can't Take a Hint

Scaling Up Liquid-Resistance Liquid-Capacitance Networks for Efficient Sequence Modeling

From n-gram to Attention: How Model Architectures Learn and Propagate Bias in Language Modeling

Analyzing Safety Risks in LLMs Fine-Tuned with Pseudo-Malicious Cyber Security Data

Defending against Indirect Prompt Injection by Instruction Detection

Using LLMs in Generating Design Rationale for Software Architecture Decisions

Evolution Meets Diffusion: Efficient Neural Architecture Generation

Direct Video-Based Spatiotemporal Deep Learning for Cattle Lameness Detection

FedDiverse: Tackling Data Heterogeneity in Federated Learning with Diversity-Driven Client Selection

Humanoid Agent via Embodied Chain-of-Action Reasoning with Multimodal Foundation Models for Zero-Shot Loco-Manipulation

Catch Me if You Search: When Contextual Web Search Results Affect the Detection of Hallucinations

COMI-LINGUA: Expert Annotated Large-Scale Dataset for Multitask NLP in Hindi-English Code-Mixing

ComfyGPT: A Self-Optimizing Multi-Agent System for Comprehensive ComfyUI Workflow Generation

CyberLLMInstruct: A Pseudo-malicious Dataset Revealing Safety-performance Trade-offs in Cyber Security LLM Fine-tuning

CoPL: Collaborative Preference Learning for Personalizing LLMs

Structured Preference Optimization for Vision-Language Long-Horizon Task Planning

A Comprehensive Survey on the Trustworthiness of Large Language Models in Healthcare

LocalEscaper: A Weakly-supervised Framework with Regional Reconstruction for Scalable Neural TSP Solvers

Forget What You Know about LLMs Evaluations -- LLMs are Like a Chameleon

Gemstones: A Model Suite for Multi-Faceted Scaling Laws

A Deep Learning Pipeline for Solid Waste Detection in Remote Sensing Images

Learning Temporal Invariance in Android Malware Detectors

Beyond checkmate: exploring the creative chokepoints in AI text

Enhancing the De-identification of Personally Identifiable Information in Educational Data

LLM-ABBA: Understanding time series via symbolic approximation

Unlocking Legal Knowledge: A Multilingual Dataset for Judicial Summarization in Switzerland

Mirror-Consistency: Harnessing Inconsistency in Majority Voting

DAVIS: Planning Agent with Knowledge Graph-Powered Inner Monologue

Semantic Alignment-Enhanced Code Translation via an LLM-Based Multi-Agent System

DPDEdit: Detail-Preserved Diffusion Models for Multimodal Fashion Image Editing

AppAgent v2: Advanced Agent for Flexible Mobile Interactions

XGen-MM (BLIP-3): A Family of Open Large Multimodal Models

Towards Unified and Adaptive Cross-Domain Collaborative Filtering via Graph Signal Processing

Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks

Database-Augmented Query Representation for Information Retrieval

Video-Language Critic: Transferable Reward Functions for Language-Conditioned Robotics

Annotation-Efficient Language Model Alignment via Diverse and Representative Response Texts

Empowering Time Series Analysis with Foundation Models: A Comprehensive Survey

FedCoSR: Personalized Federated Learning with Contrastive Shareable Representations for Label Heterogeneity in Non-IID Data

Conformal Temporal Logic Planning using Large Language Models

Learn to Relax with Large Language Models: Solving Nonlinear Combinatorial Optimization Problems via Bidirectional Coevolution

Co-Investigator AI: The Rise of Agentic AI for Smarter, Trustworthy AML Compliance Narratives

Caught in the Act: a mechanistic approach to detecting deception

When Truthful Representations Flip Under Deceptive Instructions?

TAI Scan Tool: A RAG-Based Tool With Minimalistic Input for Trustworthy AI Self-Assessment

Large Language Models' Reasoning Stalls: An Investigation into the Capabilities of Frontier Models

Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary

MAFA: A multi-agent framework for annotation

Learning Like Humans: Advancing LLM Reasoning Capabilities via Adaptive Difficulty Curriculum Learning and Expert-Guided Self-Reformulation

Designing AI-Agents with Personalities: A Psychometric Approach

Apertus: Democratizing Open and Compliant LLMs for Global Language Environments

Language models' activations linearly encode training-order recency

A Universal Banach--Bregman Framework for Stochastic Iterations: Unifying Stochastic Mirror Descent, Learning and LLM Training

Dense Video Understanding with Gated Residual Tokenization

Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting

Synthesizing Behaviorally-Grounded Reasoning Chains: A Data-Generation Framework for Personal Finance LLMs

TGPO: Tree-Guided Preference Optimization for Robust Web Agent Reinforcement Learning

Where Do Tokens Go? Understanding Pruning Behaviors in STEP at High Resolutions

Reasoning Efficiently Through Adaptive Chain-of-Thought Compression: A Self-Optimizing Framework

Queen Detection in Beehives via Environmental Sensor Fusion for Low-Power Edge Computing

Machines are more productive than humans until they aren't, and vice versa

Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices

Prompt2Auto: From Motion Prompt to Automated Control via Geometry-Invariant One-Shot Gaussian Process Learning

PhenoGnet: A Graph-Based Contrastive Learning Framework for Disease Similarity Prediction

SSL-SSAW: Self-Supervised Learning with Sigmoid Self-Attention Weighting for Question-Based Sign Language Translation

You Are What You Train: Effects of Data Composition on Training Context-aware Machine Translation Models

Hala Technical Report: Building Arabic-Centric Instruction & Translation Models at Scale

RFM-Editing: Rectified Flow Matching for Text-guided Audio Editing

MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment

Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency

Differential Privacy in Federated Learning: Mitigating Inference Attacks with Randomized Response

LLM Agents for Interactive Workflow Provenance: Reference Architecture and Evaluation Methodology

An Empirical Study on Failures in Automated Issue Solving

DSpAST: Disentangled Representations for Spatial Audio Reasoning with Large Language Models

MAP: End-to-End Autonomous Driving with Map-Assisted Planning

Caught in the Act: a mechanistic approach to detecting deception

Created by

Haebom

Author

Gerard Boxo, Ryan Socha, Daniel Yoo, Shivam Raval

Outline

This paper presents a linear probe technique that analyzes the internal activation of AI systems to detect deceptiveness in generated responses. Experiments with the Llama and Qwen models (1.5B to 14B parameters) demonstrated that, especially for larger models with 7B parameters or more, it distinguishes deceptive and non-deceptive responses with an accuracy of over 70-80%. A model fine-tuned with DeepSeek-r1 achieved an accuracy of over 90%. Layer-by-layer analysis revealed a three-stage pattern: detection accuracy was low in the early layers, peaked in the middle layers, and then slightly decreased in later layers. Furthermore, using an iterative null-space projection technique, we identified multiple linear directions that indicate deceptiveness.

Takeaways, Limitations

•

Takeaways:

◦

Analyzing the internal activation of the LLM suggests the possibility of detecting deceptive responses with high accuracy.

◦

We present a new technique that can contribute to improving the reliability and safety of large-scale language models.

◦

We present a novel approach to solving the alignment problem in AI systems.

•

Limitations:

◦

Currently, only experimental results for the Llama and Qwen models are presented, and further research is needed to determine the generalizability of these results to other models.

◦

Further analysis is needed to understand why the accuracy of deception detection varies with model size.

◦

Validation of deception detection performance in complex real-world situations is needed.

◦

A clear definition and classification of the types and extent of deceptive behavior detected is needed.

View PDF

Made with Slashpage