AI/ML arXiv cs.AI

What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness

Research exploring internal representations of LLM forecasters to improve calibration and detect when chain-of-thought reasoning is unfaithful to the model's internal state.

AI/ML arXiv cs.AI

Who Analyses the Analyser? Self-Validating LLM Hazard Analysis with Constitutional Meta-STPA

Proposal of Constitutional Meta-STPA, a framework that uses LLMs to perform safety analysis on the LLM-assisted tools themselves to create a governance constitution.

AI/ML arXiv cs.AI

Reinforcing the Generation Order of Multimodal Masked Diffusion Models

Investigation into optimizing the generation order in multimodal masked diffusion models using a learnable control module trained via GRPO.

AI/ML arXiv cs.AI

Towards Efficient Large Language Model Serving: A Survey on System-Aware KV Cache Optimization

A comprehensive survey on system-aware KV cache optimization (sKis) to reduce memory intensity and costs in LLM serving systems.

AI/ML arXiv cs.AI

When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models

Analysis of uncertainty quantification in thinking-mode Visual Language Models, finding that thinking chain entropy is a more reliable predictor than answer entropy.

AI/ML arXiv cs.AI

COBART: Controlled, Optimized, Bidirectional and Auto-Regressive Transformer for Ad Headline Generation

Introduction of COBART, a transformer-based method using prefix control tokens to generate optimized and customizable ad headlines with improved CTR.

AI/ML arXiv cs.AI

A Reliability Assessment of LALM Audio Judges for Full-Duplex Voice Agents

Researchers evaluated Gemini models as 'audio judges' for scoring full-duplex voice agents, finding that Gemini 2.5 Flash provides reliable human-like scoring across most dimensions.

AI/ML arXiv cs.AI

Who Broke the System? Failure Localization in LLM-Based Multi-Agent Systems

The AgentLocate framework was introduced to diagnose failures in LLM-based multi-agent systems by attributing errors to specific agents and the earliest decisive failure step.

AI/ML arXiv cs.AI

SpO$_2$ Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG for Oxygen Saturation Estimation

A new stage-wise time-frequency reconstruction framework improves oxygen saturation (SpO2) estimation from low-quality wearable PPG signals by integrating a predictor-guided constraint.

AI/ML arXiv cs.AI

Reaction-network reasoning with frontier models for experimentally confirmed catalyst-selectivity hypotheses

A human-AI co-thinking framework uses frontier LLMs to reason over chemical reaction networks, leading to the discovery of a copper-iron oxide catalyst with higher acetate selectivity.

Cybersecurity arXiv cs.AI

Beware What You Autocomplete: Forensic Attribution of Backdoored Code Completions

CodeTracer is a forensic framework designed to trace backdoored code completions in LLMs to the specific malicious fine-tuning data that caused them.

AI/ML arXiv cs.AI

Provably Optimal Learning Algorithms for Assistance Games

This research provides the first provably efficient learning algorithms for repeated assistance games between informed and uninformed agents, achieving optimal regret rates.

AI/ML arXiv cs.AI

Can We Trust LLM's Logic? Quantifying Uncertainty, Coherence, and Robustness via a Graph-Based Framework

GRAPHEVAL is a graph-based framework that quantifies LLM reasoning uncertainty and introduces Graph Self-Consistency (GSC) to improve reasoning fidelity over simple majority voting.

AI/ML arXiv cs.AI

APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts

APIVOT is a VLM-based planner for long-horizon robot tasks that interleaves language for semantic reasoning and visual thoughts for geometric verification.

AI/ML arXiv cs.AI

Structured Pruning of Large Language Models via Power Transformation and Sign-Preserving Score Aggregation with Adaptive Feature Retention

A new structured pruning method for LLMs uses power transformation and sign-preserving aggregation to achieve inference speedups while maintaining accuracy.

AI/ML arXiv cs.AI

DKDNet: Dual Knowledge and Data-Driven Network for Cross-Domain Automatic Modulation Classification

DKDNet is a dual knowledge and data-driven network designed to improve automatic modulation classification across different communication domains using signal prior knowledge.

Software Engineering Hacker News

In Emacs, Everything Looks Like a Service

A discussion on the architectural pattern of treating components as services within the Emacs ecosystem.

Hardware/Chips Hacker News

Tiny Tapeout Explorer: WASM FET-level circuit SIM&vis

An exploration of the Tiny Tapeout Explorer, which provides FET-level circuit simulation and visualization using WebAssembly.

AI/ML arXiv cs.AI

Efficient Safety Alignment of Language Models via Latent Personality Traits

Introduction of Latent Personality Alignment (LPA), a lightweight safety alignment method for LLMs that uses personality-anchored representations to resist jailbreaks.

AI/ML arXiv cs.AI

Adversarial Decoys: Misdirecting Attention-Based Defenses in ViT

Research on adversarial decoys in Vision Transformers (ViTs) that misdirect attention-based defenses to maintain attack effectiveness.