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

Anthropic released FlashProof for cryptographic verification of model identity

Anthropic has published FlashProof, an open-source research project for cryptographically verifying that an output comes from a specific model. The system protects against covertly swapping in cheaper models and works without revealing internal parameters. The code and technical specification are available on GitHub; deployment…

TechCrunch AI Original source ↗
Research only one source so far

Qwen 3.8 27B can interpret neural network weights without tools

The paper demonstrates that GPT-6 Astra and Qwen 3.8 27B interpret raw model weights without code. Simon Willison verified the same capabilities in Qwen 3.8 27B running on a laptop. This is Weight-Space Programming — a new approach to analyzing models.

Simon Willison — AI tag (leading independent LLM commentator) Original source ↗
Research only one source so far

Reverse Game Theory: Mechanisms for Controlling AI Agent Behavior

Researchers from MIT, Yale, and Columbia propose using reverse game theory (mechanism design) to align AI agents. Instead of analyzing the internal behavior of models, the approach focuses on designing incentives that motivate the desired behavior. Specific proposals include rewarding models for predicting the output of other…

MIT Technology Review — AI section Original source ↗
Research only one source so far

SHIFT method for unlearning in language models

A preprint presents the SHIFT method for machine unlearning, which removes specific knowledge from language models. It achieves 3.7 times better performance than existing methods, preserves knowledge best, and remains stable under relearning attacks.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

MedPsy medical language models optimized for local deployment

A preprint introduces MedPsy, a family of medical language models with 1.7 billion and 4 billion parameters designed for deployment on edge devices. MedPsy-4B outperforms MedGemma-27B while being 6.75x smaller and achieving 75% lower latency; MedPsy-1.7B outperforms MedGemma-4B despite being less than half the size. Training…

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

LLM agents can learn from video tutorials without updating parameters

A research paper introduced Tutorial Agentic Learning, a method for LLM agents to learn from video tutorials without changing parameters. VidLearn benchmark (72 chess tutorials): TAL achieved a 113% improvement on training data, 91% on test data, and opens positions better in 69% of games.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

LLM behavior in negotiations differs by language

The research examined whether LLM behavior changes depending on the language. The Gemini and Qwen models negotiated identical scenarios in English and Korean. In Korean, they achieved significantly higher success rates (77.3% vs. 52.5%). The researchers attribute the difference to more polite phrasing in Korean, particularly among sellers.

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

CULTURE-CMT benchmark for social media translation in context

A new benchmark called CULTURE-CMT contains 572 manually verified Chinese-English examples for translating social commentary. It distinguishes levels of difficulty in slang, context, and their combination. Evaluation of models shows that contextual information significantly improves translation, but larger models are not better.

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

The Legible method addresses the two-hop reasoning problem in knowledge editing in language models

The "two-hop curse" problem in language models is not a deficiency in reasoning, but a matter of how new facts are written. The Legible method ensures facts are written in a way that is legible to the model, thereby increasing knowledge composition accuracy by 1.8–2.6x over the baseline, without the need for synthetic data.

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

Structured multi-level modeling for improving language model reasoning

Researchers have introduced the Multi-Level Reasoning (MLR) method, which splits the generation of a long chain of thought into a high-level planning step and a low-level execution step. Instead of outcome-reward RL, it uses iterative Step-DPO with Twisted Sequential Monte Carlo. On mathematical and logical tasks…

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

Generalized Sparse Fuzzy Clustering Framework with Automatic Estimation of the Number of Clusters

The research paper presents the GSFC framework, which extends fuzzy clustering with sparse penalization. The algorithm autonomously estimates the number of clusters and handles noise better. It integrates a virtual sink for outliers and demonstrates better performance than conventional methods.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Research: Half of insights from analytics agents lead to improvements

MIT Technology Review study on agentic analytics in enterprises: only 48% of insights lead to improvements (37% last year), with "insight leaders" at 70%. The main problems are lack of business context (55%), fragmented data (45%), and an excess of insights (52% of companies). The solution: improving governance and semantics.

MIT Technology Review — AI section Original source ↗