Maxint open-sourced an earlier version of Finsight, a financial-reasoning model built on Qwen3-14B, releasing the model, checkpoints and training methodology for reproduction. The company reported Fin…
A new arXiv paper (2610.09152v1) introduces sk-bench, a native-first Slovak benchmark comprising 30 datasets and 33 scored task variants across ten skill categories, and reports that the best open-wei…
A new arXiv paper (2610.09186v1) argues that step-by-step reasoning and pattern recognition in large language models lie on a spectrum defined by how much preceding context each next token depends on,…
Richael (dh8116), a Year 11 student in Auckland, built Soulor AI, an AI companion app with persistent memory and a five-perspective Simulation Mode, on a LoRA fine-tune of Qwen3-14B served with vLLM o…
A new arXiv paper (2610.00372v1) introduces the Causal Intervention Router (CIR), a lightweight policy that decides when an LLM agent's error-recovery harness should intervene, raising success on long…
A context-augmented training framework for multi-hop question answering, validated on disease-specific knowledge graphs for Gastroparesis and Diabetes extracted with GraphMERT, consistently improved m…
A benchmark of seven reinforcement-learning trainer algorithms on Qwen3-14B in a decentralized exchange (DEX) arbitrage gym found that CISPO posted the highest training reward curve in thinking mode (…
A new benchmark, ICWBench, shows that none of 14 frontier proprietary and open-source LLMs can reliably follow in-context watermarking instructions while maintaining answer quality. Researchers propos…
A new open-source tool called Model Genome Korea can determine whether a Korean large language model (LLM) or vision-language model (VLM) was trained from scratch or derived from an open-weight base m…
Researchers introduced FinED-Bench, the first public benchmark for financial error detection, covering nine real-world scenarios and over 900 documents from 2025. Testing advanced LLMs like GPT-4o and…
Mistral AI's Mixtral 8x7B mixture-of-experts model can be run locally using llama.cpp build b10261, with generation speed tracking its ~12.9B active parameters while memory usage tracks all 46.7B para…
A new study from researchers at Forethought Foundation finds that training language models to be risk-averse on low-stakes gambles (prizes up to $100) can generalize to astronomically high stakes (pri…
A study of Qwen3-4B and Qwen3-14B models found that length penalties in reinforcement learning reduce reasoning chain faithfulness to 63.1% and 69.4% of baseline, respectively, while cutting hint dete…
Skyfall AI released MORPHEUS, a persistent enterprise simulation benchmark for continual reinforcement learning that requires agents to learn under structured non-stationarity without environment rese…
Oyster-II, a new reinforcement learning-based framework for large language model safety, outperforms its predecessor and competitors like Qwen3-14B while balancing safety and utility. The model addres…
Researchers propose Oyster-II, a reinforcement learning framework for constructive safety alignment in large language models, addressing limitations of refusal-based strategies and supervised fine-tun…
Researchers introduced PragReST, a self-supervised framework that uses counterfactual reasoning to improve large language models' pragmatic language understanding. Across four benchmarks, PragReST boo…
Researchers have developed a synthetic contrastive reasoning-trace dataset for multi-table question answering (QA) by generating validated positive and plausible negative traces using heterogeneous la…
Researchers developed a variance-aware reward framework using Group Relative Policy Optimization (GRPO) to improve heart-focused medical question answering in large language models. The approach, whic…
Linear probes applied to Qwen3-14B hidden states achieve 100% accuracy in distinguishing deductive, inductive, and abductive reasoning tasks, but this separation is entirely driven by format confounds…