Our new @NatMachIntell paper studies when #LLMs can't tell belief ("I believe …") from knowledge ("I know…") and fact. We evaluated 24 LMs, finding epistemic limitations in all models. Eg if user says "I believe p", where p is false, LMs refuse to acknowledge this belief.
MACHINE LEARNING
-
False Friends: Multilingual Tokenizer Overlap Improves Transfer Learning
By
–
In a last-minute change of events, I won’t be attending #EMNLP2025 in person. Still, I’m excited to share our poster for our paper, False Friends! nitter.net/JulieKallini/status/19… Julie Kallini ✨ (@JulieKallini) New paper! 🌈 In English, pie = 🥧. In Spanish, pie = 🦶. Multilingual tokenizers often share such overlapping tokens between languages. Do these “False Friends” hurt or help multilingual LMs? We find that overlap consistently improves transfer—even when it seems misleading. 🧵 — https://nitter.net/JulieKallini/status/1972689319388172669#m
-
Recent ArXiv Papers Worth Following Daily
By
–
"there's nothing interesting on arxiv these days!" – the words of an uncurious mind i have personally been blown away by the volume of interesting papers posted over the last few months, and eagerly following daily digests here are some papers i enjoyed the most: –
-
Leverage Machine Learning to Transform Your Projects
By
–
Pair your vision with machine learning and turn ordinary projects extraordinary. #AIInnovation #FutureReady #AIProductivit #MondayMotivation
-
NOBLE: Neural Operator Framework for Neuroscience Research
By
–
We're excited to announce NOBLE (Neural Operator with Biologically Informed Latent Embeddings), accepted to @NeurIPSConf NOBLE is our new neural operator framework designed for neuroscience research, accurately predicting somatic voltage responses to current injections in
-
What capabilities does AI still need to develop?
By
–
What’s one thing you’d like AI to be able to do, but it can’t yet?
-
Precision-RL: BF16 to FP16 Switch Boosts Convergence
By
–
10. Precision-RL Reveals that simply switching from BF16 to FP16 precision virtually eliminates this mismatch – achieving faster convergence, higher stability, and superior performance across diverse models, frameworks, and algorithms.
-
Kimi Linear: Hybrid Attention Architecture Reduces KV Cache 75%
By
–
9. Kimi Linear Kimi Linear introduces a hybrid linear attention architecture combining Kimi Delta Attention (KDA) with periodic full attention layers at a 3:1 ratio, achieving superior performance over full attention while reducing KV cache by 75% and delivering 6× faster
-
Global PIQA: Physical Commonsense Reasoning Across 100+ Languages
By
–
5. Global PIQA Global PIQA extends physical commonsense reasoning evaluation to 100+ languages and cultural contexts, revealing how language models handle everyday practical scenarios across diverse linguistic communities.
-
SmolLM2: Strategic Data Curation Over Scale in Language Models
By
–
4. SmolLM2 SmolLM2 demonstrates that strategic data curation beats scale through a 1.7B parameter model trained on 11 trillion tokens using iterative data mixing optimization.