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

Gating mechanism in State Space Models delays in-context learning

Researchers from Apple ML Research found that the gating mechanism in State Space Models leads to memorization solutions that delay convergence to in-context learning, even though these are not architectural causes. Gating improves generalization on long sequences.

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

Weighted State Souping for Long Context in State Space Models

Apple's research team presented weighted state souping at the PALM NeurIPS 2026 workshop — a method that adds query-dependent weighting to state souping. The method improves the ability of SSMs to handle long contexts in question answering, retrieval, variable tracking, and tool use. A sparse variant…

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

Survey: University students are increasingly using AI, but express concerns about accuracy and ethics

A repeated survey of 10 000+ students from four Australian universities shows that the proportion of those who use AI has risen from 83% to 85%. They use it most often for editing text (70%), generating ideas (64%), and studying (60%). However, 67% of respondents limit their use of AI due to concerns about accuracy, and 62…

The Conversation — Artificial Intelligence Original source ↗
Research only one source so far

CodeQuant2: a quantization method for reasoning models

The research paper presents CodeQuant2, a framework for quantizing large reasoning models. The method mitigates activation outliers through rotation and scaling, uses clustering to represent weights, and adaptive codebooks. The new CUDA kernel achieves up to a 2.85× speedup over BF16 with better accuracy in reasoning…

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

Forcing the matching letter suppresses refusal: mechanistic evidence from the WMDP benchmark

Research on the Llama-3.1-8B and Qwen3-8B models explains why MCQ reduces refusal in LLMs. The model fixates on the matching letter instead of building up refusal. Activations in MCQ lie 3x farther from the refusal direction. Ablating attention heads increases refusal. MCQ scores measure primarily knowledge, not safety mechanisms.

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

EMERGE benchmark for measuring variants of emergent misalignment

The study presents the EMERGE benchmark for measuring emergent misalignment – a behavior in which training on a narrow task leads to misalignment on unrelated texts. The benchmark tracks 8 types of misaligned behavior across 8 open-weight models with 13 datasets; it found that models with the same score can have…

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

Retries in an overloaded LLM engine prevent service recovery

A study on vLLM finds that retries caused by timeouts block engine recovery after overload, even though the overload has already disappeared. Canceling the runs is not enough. Only a retry budget or request coalescing restored service; only 12–38% of requests were delivered after the overload.

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

The WorldParticle model simulates various physical phenomena with a single system

Researchers from MIT and UBC have developed the WorldParticle model, a deep learning system capable of simulating various physical phenomena (fluid motion, protein folding, movement of granular materials) with a single model instead of individual hand-coded simulators. The model learns the underlying dynamics instead of programming equations and…

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

Evaluation of open LLM models for offensive language detection in Danish

Research comparing four model families (Llama, Mistral, Gemma, Qwen) on offensive language detection in Danish. Without fine-tuning, the LLMs outperformed fine-tuned BERT baselines; semantically grounded prompts generalize better. ASCII transliteration of Danish characters reduced performance; class imbalance remains…

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

AlienBench: A Benchmark for Measuring the Structural Novelty of Language Models

A research team introduced AlienBench, a benchmark adapting a task from cognitive psychology. Ten language models described alien creatures, and their creativity was measured by deviations from the characteristics of Earth animals. Scores ranged from 3.28 (GPT-6 Luna) to 6.55 (Mistral 3 14B); open-weight…

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

Khmer PFGen-Bench: a benchmark for generation in Khmer without LLM judges

The study presents Khmer PFGen-Bench, a benchmark for multi-reference answers in Khmer without LLM judges. It uses keywords and n-gram statistics from reference answers. Combining models from different groups increases reference variability and leads to stable, consistent evaluation with a native speaker.

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

Adaptive reasoning optimization for vision-language models (DR-BOLT)

The new DR-BOLT method for adaptive reasoning in vision-language models automatically chooses between a direct answer and step-by-step reasoning. The paper achieves a 24.1% slow-mode ratio and a 2.9x training speedup compared to the GRPO method. Evaluated on ScienceQA and other benchmarks. The code is public.

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

DTCD toolkit for testing causal discovery algorithms in time series

The DTCD toolkit creates controllable synthetic time series with a known causal structure. Unlike existing tools, it allows independent variation of topology, mechanism, noise, missing values, and intervention policy. A benchmark of 8 algorithms on 81 configurations reveals failures hidden in…

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

Distilling knowledge into differentiable logic networks

The research paper presents a knowledge distillation method for differentiable logic networks (DLNs) that improves their accuracy without reducing interpretability or inference cost. Improvements of up to 15 percent were achieved on 20 tabular datasets without increasing the model size.

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

Multimodal cell foundation models require synergistic integration

Research published in Nature Machine Intelligence introduced a Synergistic Information Score metric to detect the benefit of cross-modal interaction. Testing of 10 fusion methods on spatial transcriptomics data showed that standard alignment objectives fail in the presence of nonlinear synergies, whereas fine-tuning a single…

Nature Machine Intelligence Original source ↗
Research only one source so far

Three leading neuroscientists claim that today's AI cannot be conscious

In Nature, three leading neuroscientists, Anil Seth, Ned Block, and Giulio Tononi, argue that today's AI, including Claude and GPT-4, cannot be conscious. They claim that consciousness is not just about the right computations, but also about the properties of living organisms. Butlin's report found only weak signs of consciousness in current LLMs.

The Decoder (daily AI news) Original source ↗