All Articles
16483 articles total
The Devil is in the Spectrum: Mitigating Representation Collapse in LLMs via Topologically Regularized Side-Path
A research paper proposing TRSP to mitigate representation collapse in LLMs by regularizing token interaction topology.
Expectation Alignment of Language Models for Real-World User Expectations
Presentation of ExpectBench, a benchmark for evaluating LLM alignment with real-world user expectations, and the LENS framework for better alignment.
OPTScientist: Multi-Agent Discovery of Typed Optimizer Programs for Transformer Pretraining
A multi-agent framework called OPTScientist for the automated discovery of typed optimizer programs for Transformer pretraining.
Directional Hallucinations: Ideological Drift in News-Grounded LLM Question Answering
An investigation into ideological drift and hallucinations in LLMs when answering questions based on news articles.
Autonomous Topology Mutation: Safe Runtime Restructuring for Multi-Agent LLM Systems with Capability, State, and Shadow Invariants
Introduction of ATM, a mechanism for multi-agent LLM systems to perform safe, runtime topology mutation to handle agent overload.
EvoSQL: Memory-Augmented Critic-Generator Co-Evolution for Text-to-SQL
A co-evolution framework called EvoSQL for improving Text-to-SQL tasks through memory-augmented critic-generator interaction and SDPO fine-tuning.
Quickly rewrite Git repository history
A discussion on techniques for quickly rewriting Git repository history.
Why Sony can't bring back its classic Walkman models
An analysis of the challenges Sony faces in reviving classic Walkman models.
Robust Critics: Defending LLMs Against Multi-Turn Attacks
Proposes Dialogue Critic Guided Sampling (DCGS) to defend LLMs against multi-turn attacks by inferring user intent throughout a conversation.
Incomplete Prompt Jailbreaks in Large Language Models
Investigates incomplete prompt jailbreaks in LLMs and identifies specific functional neurons (termination and continuation) for potential defense interventions.
VeriSimpl: Robust Optimization Modeling from Natural Language using Simplification-based Verification
Introduces VeriSimpl, a framework that uses simplification-based verification to improve the translation of natural language into robust optimization models.
SonicSampler: Unified Tile-Aware Kernels for LLM Sampling and Speculative Verification
Presents SonicSampler, a suite of Triton kernels that vertically fuse the LLM sampling pipeline, achieving up to 16x speedup in speculative verification.
Benchmarking Large Language Models on Multi-Sensor Physical Hazard Assessment
Benchmarking study showing that LLMs fail to identify physical hazards when multiple sensors are slightly elevated but below individual safety thresholds.
Semi-Supervised Text-Attributed Graph Distillation
Proposes a semi-supervised framework using Wasserstein Distance for efficient distillation of Text-Attributed Graphs (TAGs).
Beyond Liars' Bench: The Impact of Lie Typology, Depth, and Sparsity on Deception Detection in LLMs
Analyzes how lie typology and representation depth affect the ability of probes to detect deceptive outputs in LLMs.
Enabling Scalable Topology Inference in Distribution Systems via Constrained Multi-Source Inference
Presents a constrained multi-source inference framework for scalable and accurate topology recovery in electrical distribution systems.
AINTMA: Agentic AI Architecture for Autonomous Test Management with Generative Intelligence, Secure Cloud Communication and Adaptive Quality Analytics
AINTMA is a multi-agent AI system designed for autonomous software test management, utilizing six specialized agents to reduce test cycle time and defect escape rates.
Marking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts
Research indicates that LLM watermarking can degrade medical reasoning quality and introduce hallucinations, highlighting the need for domain-specific evaluation in healthcare.
ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models
ClickGuard is an AI-driven browser extension that detects clickbait using a hybrid ML architecture and provides users with summaries of the content.
Stochastic Sampling is Epistemically Shallow: The Dimensionality Gap Between Temperature Variation and Model Diversity in LLMs
A study reveals that stochastic sampling in LLMs is 'epistemically shallow,' meaning it provides per-question uncertainty but lacks the cross-question structure found in diverse ensembles.