Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues
LLMs exhibit performative compliance: fairness evaluations overestimate moral safety when demographic identity must be inferred rather than labeled.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
LLMs exhibit performative compliance: fairness evaluations overestimate moral safety when demographic identity must be inferred rather than labeled.
Bayesian workflow calibration detects and repairs statistical errors in probabilistic programs written by LLMs using posterior checks and diagnostics.
Team MKC applies LLMs to detect mental health changes from social media timelines for early detection and monitoring.
Uncertainty-aware decision-making algorithms for LLMs based on Bayesian theory to improve reliability in ambiguous tasks.
Poller uses LLMs as poetry evaluators by role-playing authors for Chinese poetry understanding tasks.
DeepReinforce releases Ornith-1.0, MIT-licensed open-weights model (9B–397B variants) for agentic coding, built on Gemma 4 and Qwen 3.5, achieving SOTA on coding benchmarks.
Argues LLMs lack situated perception of physical world causality and agency; identifies this primitive as necessary prerequisite for artificial superintelligence.
Study compares nine open-weight LLMs against human participants in networked Prisoner's Dilemma, finding selected models reproduce cooperation macro-dynamics but lack individual fidelity.
Framework translates natural language system requirements to ATL/A formal specifications for multi-agent systems verification using LLMs.
PromptGNN-sim: bidirectional GNN-LLM fusion framework for text-attributed graph learning via Graph Attention Networks and semantic-structure alignment.
Multi-agent system using open-source LLMs for automated disinformation detection and fact-checking at scale across social media.
Empirical study across 25 LLMs showing chain-of-thought gains derive from semantic content, not token length, using in-distribution paired comparisons.
Conceptual analysis: LLMs as degenerate world models; latent-space architectures as strict generalization of autoregressive token prediction.
LLMs applied to data fusion/truth discovery in tabular data via prompting strategies for single and multi-truth scenarios.
Study shows LLM self-evaluation is not uniformly easier than generation; multi-hop QA reveals attention-based mechanistic differences.
Extractive-abstractive hybrid summarization for legal case judgements using tree-of-thoughts with DeepSeek and Llama.
RiVER framework enables RL training of LLMs on optimization tasks without ground-truth answers using continuous execution feedback.
Study quantifies alignment between sequence probability and correctness across LLM decoding methods, models, and benchmarks.
Language-based digital twin framework using LLMs detects Mild Cognitive Impairment via stylometric analysis and conversation patterns.
PEEU method trains open-source MLLMs for GUI task planning via autonomous environment exploration and hindsight experience replay.
Training-free source selection for LLMs using Fisher alignment metrics at vocabulary scale for scientific domains (SMILES, proteins).
Sparse autoencoders reveal LLMs isolate time-aware vs. look-ahead-biased features; feature steering improves cross-domain forecasting generalization.
TOPS framework prunes visual tokens in MLLMs via principled optimization, balancing instruction-relevance and diversity for faster inference.
Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo…
Standard benchmarks for multimodal large language models (MLLMs) score each item on one canonical ordering and miss whether order-irrelevant shuffling changes the answer, a baseline reliability property called for by emerging AI evaluation guidelines. We introduce Facet-Probe…
Training-time data poisoning during fine-tuning poses a significant threat to large language models (LLMs) deployed for abstractive text summarization, where small task-specific datasets exert disproportionate influence on model behavior. In this setting, adversaries manipulate…
Multimodal Large Language Models (MLLMs) demonstrate strong performance on standard visual question answering benchmarks, yet their scalability under controlled structural complexity remains poorly understood. We introduce TriViewBench, a controlled three-view visual reasoning…
Tool use enables large language models (LLMs) to perform complex tasks, and recent agentic reinforcement learning (RL) methods show promise for enhancing model capabilities. However, RL alone often leads to instability or limited gains in tool-use tasks. In our experiments…
Text detoxification, the automated detection and mitigation of abusive and harmful content, is essential for ensuring the safety of online communities and protecting users. However, low resource languages such as Tatar have received little research attention. In this paper we…
Large language models (LLMs) attain remarkable surface fluency on code, yet they neither formally guarantee the syntactic validity of their output nor leverage the hierarchical structure defining the target language. While existing constrained-decoding frameworks address the…