Asymptotic Semantic Collapse in Hierarchical Optimization
• Asymptotic Semantic Collapse: dominant context absorbs individual semantics in multi‑agent language systems. • Dominant Anchor Node with infinite inertia drives asymptotic alignm
Nlp
ConfSpec: Efficient Step-Level Speculative Reasoning via Confidence-Gated Verification
• ConfSpec introduces confidence‑gated cascaded verification for step‑level speculative reasoning efficiently. • Small draft models quickly verify reasoning steps, accepting high‑c
Emergent Dark Patterns in AI-Generated User Interfaces
• AI-driven UI design now adapts and personalizes, but also amplifies dark pattern risks. • Dark patterns exploit psychological biases, learned from existing deceptive data, becomi
Exploring the Ethical Concerns in User Reviews of Mental Health Apps using Topic Modeling and Sentiment Analysis
• AI‑driven mental‑health apps raise ethical concerns about privacy, bias, and user trust. • Researchers built an NLP framework to analyze user reviews from Google Play and Apple A
INSURE-Dial: A Phase-Aware Conversational Dataset \& Benchmark for Compliance Verification and Phase Detection
• INSURE‑Dial is the first public benchmark for compliance‑aware voice agents in insurance calls. • Corpus contains 50 real AI‑initiated calls and 1,000 synthetic calls, averaging
Prompt Optimization Via Diffusion Language Models
• Diffusion-based framework refines system prompts via masked denoising in an iterative manner. • Conditions on interaction traces: user queries, model responses, and optional feed
Spilled Energy in Large Language Models
• Reinterprets LLM softmax as Energy-Based Model, enabling energy tracking during decoding. • Introduces training‑free metrics: spilled energy and marginalized energy from logits.
When large language models are reliable for judging empathic communication
• LLMs generate empathic responses, but reliability of judging empathy remains unclear. • Study compares expert, crowdworker, and LLM annotations across four psychological framewor
DeepContext: Stateful Real-Time Detection of Multi-Turn Adversarial Intent Drift in LLMs
• DeepContext introduces stateful monitoring for LLM safety, tracking intent across turns. • Uses RNN to process fine‑tuned turn‑level embeddings, preserving conversation context.
KD4MT: A Survey of Knowledge Distillation for Machine Translation
• KD used for compression and knowledge transfer in MT, shaping supervision and translation quality. • Survey covers 105 papers up to Oct 2025, providing comprehensive landscape of
Differential Transformer V2
• DiffTransformer V2 doubles query heads, keeps KV heads constant for efficient attention. • Uses differential attention: subtracts paired heads within same GQA group. • Applies si
Search Query Understanding with LLMs: From Ideation to Production
• Yelp integrates LLMs to interpret search queries, improving intent detection for millions of daily searches. • The team tackled spelling correction, segmentation, canonicalizatio
Language, Statistics, & Category Theory, Part 3
• Introduces enriched category theory framework for modeling language expressions and their relationships. • Builds on Part 2’s set assignment to words, extending to statistical co