Skip to content

Research separate from news

What could become important next

A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.

1381 published research events

Latest work

A significant claim from a single source is published only after further confirmation.

Research only one source so far

Valence and morality in natural language: a pilot annotation study

A research study examining the relationship between the emotional tone (valence) of text and morality. The team created a dataset of 500 annotations from six people for the moral valence of actions and consequences, ranging from -1 to 1. A regularized logistic regression model achieved Matthew's correlation coefficient 0.764 for binary classification. The results suggest that valence is useful for estimating the morality of text.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Mixture-of-Experts routing follows Huffman coding principles, study shows

Researchers discovered that routing in Mixture-of-Experts models (Phi-3.5-MoE, Gemma-4-27B-A4B, Qwen3.5-35B-A3B) follows Huffman coding principles — allocating sparse expert resources to common tokens and more diverse expert committees to rare tasks. They propose Subset Difference Pruning to…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

CAMeR: Adaptive memory in LLM agents with hybrid word and embedding gating

Research introduces CAMeR, a memory system for LLM agents that combines gating at the word level (Jaccard) and the embedding level (cosine similarity) with adaptive weights. CAMeR-Bench, a new benchmark (76 memories, 100 rounds) in 8 thematic groups, shows 1.6× greater discrimination between frequently used and never…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Detecting user interface principle violations with reinforcement learning

A research team in an arXiv study trained a 4B vision-language model to detect violations of 19 UI quality principles (WCAG 2.2, deceptive design, perception/cognition) on a synthetic dataset of 10 thousand pages. The model achieved 84% micro-F1 (up from 36%); 13 of the 19 principles exceeded 80% F1. The resulting critic can…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Auditing evaluation–deployment mismatch in fine-tuned language models

A research paper describes a method for detecting differences between model behavior during testing and ordinary deployment. The authors locate internal representations of these differences and test editing model behavior. The method works in 10 of 12 cases, but does not guarantee deployment safety.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study of diversity in LLM opinions: Persona conditioning more effective than simple methods

A scientific study tests interventions to increase diversity in LLM opinions. It found that simply increasing the temperature and providing instructions are ineffective; structured persona conditioning combined with interaction architectures is effective. The results are based on a factorial experiment with 100 questions and 7 models.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Detecting toxicity in gaming chats using a neuro-symbolic approach

The research paper describes a system for classifying toxicity in gaming chats. It combines an ensemble of compact transformers (DeBERTa-v3-base, XLM-RoBERTa-base) with a linguistically informed mediator. The system ranked 3rd in Macro F1 and 1st in accuracy among participants in the EEUCA 2026 competition. It addresses…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Protecting language models from knowledge distillation by editing logical schemas

A research preprint proposes the SGRE method for protecting LLMs from knowledge distillation by modifying reasoning chains in responses. The method preserves accuracy and naturalness of the text while making them harder to copy. Tests demonstrated reduced distillation effectiveness without a loss of quality.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Shared-private fusion of reward models for preference evaluation

A research paper proposes CSPF to improve evaluation of non-verifiable tasks by combining multiple reward models. The method decomposes evaluator signals into shared and specific representations under human preference supervision. On the LM-Arena and PPE datasets, it achieved the best results compared with…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Instruct-FD: How well full-duplex dialogue systems follow instructions for speech control

The new Instruct-FD benchmark tests how well full-duplex spoken dialogue systems follow speech-control instructions. A study of six state-of-the-art systems found that even the best model follows only 64.4% of instructions. They particularly struggle with proactive behaviors such as backchanneling and interruption.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Learn2Zinc: Fine-tuning small language models to translate natural language into MiniZinc

A research team examines how to teach small language models (0.6–20B parameters) to generate MiniZinc code from verbal descriptions. They found that syntax errors dominate failures; they propose collecting errors from multiple runs and using them to train repairs. With fine-tuning, they achieved 98% execution accuracy…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Measuring bias in LLM world models using prediction markets

Research shows that LLMs are biased by the text in the corpus they were trained on. The authors propose a methodology for measuring bias using prediction markets as an external reference. In an experiment on 111 Ukrainian markets (93 000 predictions, four models), they found that English-language news coverage systematically introduces bias…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

THOR: A hierarchical oscillatory framework for multi-hop question answering

The THOR research framework addresses two key problems in multi-hop question answering: attention decay during long reasoning chains and error accumulation. Inspired by theta-gamma oscillations in the brain, it separates global planning from local retrieval and introduces a verification mechanism to stop…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

REGARD: Regional differences in the emotional framing of large language models

An arXiv study examines how 19 large language models frame political and cultural entities differently in the post-Soviet region. Using Valence-Arousal-Dominance profiling, it identified three model clusters with different affective behavior, with the rate of generic responses strongly correlated with lower…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

From agent failures to textual policies: what works and what fails

Research results on TextGrad applied to agents. Handwritten policies improved the success of frozen 7B agents by 5.0 points, but policies generated from data performed no better than fixed prompting. The main challenge is reliably generating and selecting policies from experience, rather than executing them.

arXiv cs.CL (Computation and Language / NLP) Original source ↗