Natural Language Processing

Researchers have developed TPA (Think, Plan, Ask), a proactive multi-agent dialogue framework that enables large language models to systematically select questioning strategies to surface latent social language disorder …

A new study of small language models reveals that chain-of-thought prompting for arithmetic relies on a positional shortcut: the model copies whichever number appears last before the answer delimiter, regardless of the l…

Researchers have introduced Transcoders, a function-centric framework that decomposes vision-language models into interpretable computational pathways linking image patches to token generation. Applied to Gemma 3-4B-IT, …

Researchers introduced HawkesLLM, a framework that models how uncertainty propagates through sequential text generation in agentic simulation systems. The framework uses a multivariate Hawkes process to track temporal in…

Researchers decomposed GPT-2 XL and Llama-3.1-8B into 16,000 to 32,000 interpretable features per layer using sparse autoencoders, finding that semantic features alone recovered 94% of peak brain-response encoding perfor…

A study of fMRI data from 112 English, Chinese, and French speakers found that brain-LLM alignment is driven by training-language dominance, not an inherent property of English. A Chinese-dominant model reversed the alig…

Researchers have developed a reproducible NLP pipeline for Katharevousa Greek, a historical language variant used in Greek parliamentary archives, by creating a Universal Dependencies-style parsing resource from 1,697 se…

Researchers developed a graph neural network that models alignment topology between reference documents and LLM outputs to detect hallucinations, achieving state-of-the-art results on four datasets. The method outperform…

Researchers introduced NeuroNL2LTL, a neurosymbolic framework that translates natural language into Linear Temporal Logic (LTL) with formal correctness guarantees. The system uses a structure-preserving intermediate repr…

Researchers introduced PathCal, a training-free decoding controller that calibrates reasoning paths in large language models by distinguishing distinct types of reflection markers such as "wait" and "but." The system int…

Researchers have identified a fundamental flaw in how AI agents are evaluated for knowledge work, finding that higher benchmark scores do not reliably indicate real-world performance. The team proposes a three-step frame…

Researchers introduced parallel context compaction for long-horizon LLM agent serving, addressing the problem of growing conversation histories exceeding context windows. The method provides fine-grained, predictable con…

A new study evaluating large language models in the social deduction game Secret Hitler found that current architectures remain ineffective at complex, multi-turn manipulation and deception. Models like Llama 3.1 70B ach…

Researchers have developed Query-Adaptive Semantic Chunking (QASC), a method that integrates user queries into the document segmentation process for Retrieval-Augmented Generation systems. In tests on 100 technical docum…

Researchers have developed a knowledge-aware Text-to-SQL framework that improves the performance of large language models in converting natural language questions into executable database queries, particularly in low-res…

Researchers introduced Reflection-Augmented Scaling (RAS), a method that uses execution error messages from databases to improve Cypher query generation through in-context learning. Testing across three Neo4j datasets an…

Researchers have developed MedExpMem, an experience memory framework that enables medical vision-language models to accumulate differential diagnosis expertise from their own diagnostic failures. The system, which mirror…

Researchers have developed LIFT, a new fine-tuning algorithm for diffusion language models that improves reasoning capabilities by aligning token learning difficulty with the information available at different diffusion …

Researchers tested seven large language models on their ability to infer employee expertise from 27,188 Slack messages, finding that Gemini 2.5 Flash achieved the lowest estimation error at 21.13% mean absolute error. Th…

A new study found that large language models (LLMs) exhibit systematic and reproducible biases when advising on religious conversions, consistently favoring some faiths while discouraging others. Researchers tested 20 mo…