AI
Researchers have built a method to make artificial intelligence more reliable when it reads emotions in text, such as clinical notes, counselling chats and posts in online support groups. This matters because early triage and risk assessment often depend on what people write and how that writing is interpreted. Why
AI
Researchers have a new explanation for why some AI systems get better without human feedback. The work matters because it may cut the need for expensive labeled data, but also because self-improving systems can lock in their own mistakes if left unchecked. Why this is being discussed now A team
AI
Not all fine-tuned LLMs spot phishing equally. A new study tests Llama 3.1 8B, Gemma 2 9B, and Mistral on high-stakes phishing detection—and uses SHAP and mechanistic interpretability to reveal why models do (or don’t) generalize. * Architecture × data diversity matters: Gemma 2 9B hits state-of-the-art performance (F1
AI
Jailbreak prompts can still push chatbots to produce harmful content, even after safety training. Many defenses either miss sophisticated attacks or over-block harmless requests. This paper finds a useful clue: during generation, models carry latent safety signals that spike when content turns risky, but those signals get overridden by the
AI
AI that reads privacy policies so you don't have to Privacy policies are long, legal, and hard to compare. LADFA is a new AI framework that turns those walls of text into a clear map of how companies collect, use, store, and share your personal data. How it
AI
New AI that keeps creativity: DPWriter LLMs trained with reinforcement learning often play it safe, shrinking the variety of their stories. DPWriter flips the script by planning before writing. It breaks generation into semi-structured steps, then uses Diverse Planning Branching to explore multiple, intentionally different routes. A group-aware diversity reward
AI
Large language models can ace math benchmarks yet still stumble on simple number sense because they treat numbers like ordinary words. This work fixes that by giving models a value-aware way to read numbers. How it works: whenever a number appears, the input is augmented with a tiny prefix token
AI
TL;DR An LLM acts as a coach to prune another LLM, shrinking it ~45% while preserving key knowledge and accuracy. Traditional pruning uses fixed rules and often wipes out facts. This paper lets a foundation model adaptively choose which layers to trim each round. It reads layer sensitivity snapshots—
AI
TL;DR Debiasing big language models is costly. This study shows compact “BabyLMs” can mimic how larger BERT-style models learn biases—so researchers can test ideas faster and cheaper. * BabyLMs (small, BERT-like models on tiny, editable corpora) track the same bias and performance patterns as standard BERTs. * Correlations hold across
AI
Leaderboards drive Text-to-SQL progress—but what if the test sets are wrong? This study audits two popular benchmarks and finds widespread annotation errors that can flip who looks best. * Error rates: 52.8% in BIRD Mini-Dev, 62.8% in Spider 2.0-Snow. * After correcting a BIRD Dev subset, open-source agents
LLMs
Can we make peer review fairer by rating reviewers like chess players? This study simulates a conference where multiple LLM agent reviewers with distinct personas evaluate papers across several rounds, guided by an Area Chair (AC). Researchers compared a baseline setup to versions that add Elo ratings (to track reviewer
AI
AI is shaping what we read and decide. But do large language models carry hidden political leanings? A new study proposes a transparent way to audit political bias: ask models to predict how parliaments voted, then compare those predictions to the real records. * Benchmarks: PoliBiasNL (2,701 Dutch motions; 15