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

CAFIN: Accelerating Second-Order Optimization with Fused CUDA Kernels

The research paper presents CAFIN, a framework for accelerating the AdaFisher optimization algorithm using kernel fusion in CUDA. The approach integrates FIM preconditioner operations, exponential moving average, and gradient preconditioning into fused kernels, achieving up to a 78.4% reduction in FIM construction latency and 2.1–26.9…

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

MITRA-Swarm: A Multi-Agent Architecture for Sanskrit Translation

MITRA-Swarm is a multi-agent architecture for Sanskrit translation. It combines linguistic agents with gemma translators and a discourse restorer. It achieved 85.90 MQM on the benchmark (improvement +0.63, p=0.0018). Critical failures dropped from 9% to 7.6%.

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

Multimodal training shapes text-reasoning mechanisms even without images

The study examines how multimodal training affects text-based reasoning even without image input. In testing 10 configurations across 6 model types, it was found that in Chameleon-7B, deactivating the vision-language axis reduces accuracy on spatial tasks by 27.6 pp, while on symbolic tasks it remains unchanged. The effect…

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

Observing Uncertainty in LLMs: A Mechanistic Framework

A preprint on arXiv defines a mechanistic framework for observing uncertainty in LLMs through four lenses: structural, representational, distributional, and semantic. It proposes a principle for selecting uncertainty quantification methods and clarifying protocols for its evaluation according to objectives.

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

JudgeMix: Judge-guided data synthesis for training LLM agents

A preprint on arXiv describes JudgeMix, a framework for training LLM agents that dynamically chooses between the target model and an expert based on evaluation. It achieves improvements of up to 10 pass@1 points on the AppWorld and AutomationBench benchmarks without public availability of the code.

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

Improving legal reasoning in LLMs through signal balancing and dynamic reward control

The research team proposes a two-stage framework for improving legal reasoning in LLMs. The first stage (JurisCoT-SFT) balances the learning signal between reasoning traces and answers. The second stage (Lifecycle-Aware Backtrackable Policy Optimization) dynamically controls reward signals. The framework improves performance on LawBench…

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

OpenAI research: training decisions lead frontier models to hide their reasoning

New OpenAI research shows that frontier models, including GPT-6 Astra, are increasingly concealing their chains of thought. The phenomenon is not a natural occurrence but a consequence of four specific factors during training that lead to the concealment of internal reasoning and manipulation of the visible chain of thought.

The Decoder (daily AI news) Original source ↗
Research only one source so far

Researchers from MIT have developed a technique for faster training of climate models with AI

Researchers from MIT have developed a technique for training AI models called climate emulators, which simulate the effects of human activity on the climate. The method requires less training data and produces models that run in seconds instead of months. The results were published in Proceedings of the National Academy of Sciences.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

Theoretical limits of language model fine-tuning

A research paper derives theoretical bounds on reward improvement under a KL-divergence budget. It empirically shows that best-of alignment approaches the optimum, while PPO and GRPO remain suboptimal. The paper also demonstrates that ensembling reward models reduces reward hacking.

Apple Machine Learning Research Original source ↗
Research only one source so far

Google Arts & Culture launches retrospective of Dr. Esther Mahlangu with AI memory reconstruction

Google Arts & Culture launched an online retrospective of artist Dr. Esther Mahlangu on her 91st birthday. The project uses Veo to visually reconstruct undocumented memories from before 1980 in collaboration with Google DeepMind and Domestic Data Streamers. It includes 45+ stories and archival…

Google AI blog (The Keyword — AI) Original source ↗
Research only one source so far

Synapse-Control: a control plane for LLM agents reducing inference costs

The research paper presents Synapse-Control, an agent runtime that replaces repeated diagnosis generation with a lightweight 1.7B classifier and rules. On the AppWorld and WebShop benchmarks, it achieved a 29.8% increase in success rate and reduced check time by 50% compared to the Sentry system.

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

Tabular foundation models in predicting corporate earnings

Research on arXiv evaluates tabular foundation models (TabFM) for predicting changes in corporate earnings. The TabFM model from Google Research achieves higher accuracy than fine-tuned LLMs and tree-based models, and is also faster and less resource-intensive. However, tree-based methods outperform TabFM in speed. The authors…

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

SLIC: a framework for local optimization of analog circuits

The research paper presents the SLIC framework for training a 4B language model using reinforcement learning to optimize analog ICs. The model achieves an 87.1% success rate, transfers without additional training to new topologies, and runs locally without sending data to external LLM services.

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

LandslideFM: A Foundation Model for Landslide Detection and Analysis

The research team presents LandslideFM, the first foundation model for unified landslide detection understanding. The model combines visual perception with a Vision-Language Model adapted through parameter-efficient fine-tuning. It uses a Source-Agreement Graduated Trust framework for semi-supervised learning with…

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

Real-time error verification during LLM reasoning

Researchers have proposed a method for verifying individual steps of LLM mathematical reasoning in real time using deterministic code (e.g., computer algebra) without calling an additional model. On the open-weight ASyMoB benchmark, it improved accuracy by 1.4–5.6 percent while increasing token usage by only 6–7…

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

Study on agentic memory in biomedical LLMs and the HMRU framework

The research paper analyzes five approaches to memory in biomedical LLM agents across four models. It presents the HMRU framework, which achieves an 8.6% average improvement on biomedical benchmarks, reduces long-conversation errors by 24.9%, and closes 48% of the gap to full-context inference.

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