Joshua Kazdan

@Piotr761303Ueh @jchudnov @RylanSchaeffer @sanmikoyejo @stai_research We start to see identification of semantically equivalent pairs around 3-4B param models. We are not sure at what point it begins to have a negative impact on training.

@jchudnov @JoshuaK92829 @RylanSchaeffer @sanmikoyejo @stai_research Thanks! Is 3-4B range considered a larger model in this context?

Your deduplication pipeline was built for small models. At scale, it's broken. New preprint: "Scale Dependent Data Duplication" 1/10

@AlexanderSpangh @jchudnov yes! If you take a look at Fig 2 that's exactly what it shows. The longer you train the model, the more the gradients induced by semantically identical documents align.

Super cool work! You mention "stronger models encode data better", but in your plots I mainly see # model parameters and model size. Does this also imply that the same-size model trained longer, better etc. will also make less effective use of its dataset? Does data effectiveness decrease throughout training?

If you train large models, curate pretraining data, or care about whether scaling laws actually hold, this is for you. Preprint: arxiv.org/abs/2603.06603 With @JoshuaK92829, Noam Levi, @RylanSchaeffer, Abhay, Bo, Mehmet, @sanmikoyejo, and David Donoho @stai_research @StanfordAILab @stanfordnlp 10/10

Here's hoping for better luck at ICLR 2026! openreview.net/forum?id=QzIQg… If you want to read the paper without R7Hk's endorsement: arxiv.org/abs/2502.19537 @DjDvij also made a colab where you can try the attack out for yourself: colab.research.google.com/drive/1FLbE9VP…

3. Writing the majority of your review using a language model. It did such a great job! Thanks also to the AC for ignoring us when we reported this review for violating the @NeurIPSConf guidelines against LM reviewing.

So exuberant to announce that our paper "No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanisms" has been rejected from NeurIPS 2025 with an average score of 4! 💪🔥🔥💯 @DjDvij @RylanSchaeffer @sanmikoyejo @ChrisCundy @AbhayPuri98

New position paper! Machine Learning Conferences Should Establish a “Refutations and Critiques” Track Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas 1/6

@casper_hansen_ @RylanSchaeffer There's no contradiction. We don't claim that min-p is better or worse than other logit processors-- we contend only that the evidence in Minh et. al. does not meet scientific standards to claim superiority.

@RylanSchaeffer I use min_p and it improves coherence in my experience. The claims made here are in direct conflict with my and other people’s experience

🚨New preprint 🚨 Turning Down the Heat: A Critical Analysis of Min-p Sampling in Language Models We examine min-p sampling (ICLR 2025 oral) & find significant problems in all 4 lines of evidence: human eval, NLP evals, LLM-as-judge evals, community adoption claims 1/8

A bit late to the party, but our paper on predictable inference-time / test-time scaling was accepted to #icml2025 🎉🎉🎉 TLDR: Best of N was shown to exhibit power (polynomial) law scaling (left), but maths suggest one should expect exponential scaling (center). We show how to ... 1/3

Interested in test time / inference scaling laws? Then check out our newest preprint!! 📉 How Do Large Language Monkeys Get Their Power (Laws)? 📉 arxiv.org/abs/2502.17578 w/ @JoshuaK92829 @sanmikoyejo @Azaliamirh @jplhughes @jordanjuravsky @sprice354_ @aengus_lynch1 @_robertkirk

🚨 New Paper 🚨 An Overview of Large Language Models for Statisticians 📝: arxiv.org/abs/2502.17814 - Dual perspectives on Statistics ➕ LLMs: Stat for LLM & LLM for Stat - Stat for LLM: How statistical methods can improve LLM uncertainty quantification, interpretability, trustworthiness & more. - LLM for Stat: How LLMs can enhance statistical workflows: from data collection, synthesis, annotation to statistical modeling, with applications to medical research Presents key LLM advances: Architecture, Training, Reasoning, and Self-Alignment: (1) 🧠Evolution of LLM architectures with Transformers and Self-Attention (2) LLM training pipeline from pre-training, SFT, to RLHF and Preference Optimization. (3) 💭 System 2 Prompting and Chain-of-Thought for test-time scaling . (4) 🚀 LLM Self-Alignment for achieving super-human intelligence Statisticians play a key role in the development of large-scale AI models: (1) 💡 Statistical insights improve LLM uncertainty quantification & interpretability (2) 🤖 Watermarking for AI-generated content detection (3) ⚖️ Privacy & algorithmic fairness to ensure responsible AI adoption LLMs can also empower statistical science by: (1) 📈 Scaling up data collection, synthesis, and annotation. (2) 🖥️ Automating statistical coding & exploratory analysis (3) 🔬 Facilitating medical research By bridging statistics & AI, we can: ✅ Improve better LLMs with statistical methodologies. ✅ Leverage LLMs for statistical applications in high-stakes domains

(1/n) Fine tuning APIs create significant security vulnerabilities, breaking alignment in frontier models for under $100! Introducing NOICE, a fine-tuning attack that requires just 1000 training examples to remove model safeguards. The strangest part: we use ONLY harmless data.

@arundsharma @belindmo @KyssenYu @proudmpala @sanmikoyejo @pydantic Thanks for bringing this up-- I'm surprised to hear you got such a low accuracy. We're happy to share our evals. Let's connect over email?

@belindmo @JoshuaK92829 @KyssenYu @proudmpala @sanmikoyejo FWIW, I got 20.06% accuracy with gemini-2.0-flash and qwen2.5 (localhost, open weights model) as the evaluator. About 15 essays had erroneous answers since gemini generated answers that didn't pass @pydantic.

New package + paper drop 📄 - Introducing KGGen – a simple library to transform unstructured text into knowledge graphs. Text is abundant, but good knowledge graphs are scarce. Feed it raw text, and KGGen generates a structured network of entities and relationships. (1/7)