Tag: text
Call the Science Police
2026-05-25 alignment sampling text toxicity Proposal to improve the scientific accuracy of LLM output in domains like medicine, by using a larger model to write executable rules that are applied to a smaller model's output at search time. Access: $ Basic
LLMs are Predictably Predictable
2026-05-18 alignment evaluation sampling text Some tasks we'd like language models to do, require them to make random selections. Are the models actually able to do that without external help? Access: $ Basic
Dissociated Press
2026-05-11 basics model-intro text Dissociated Press is a very simple non-neural-net language model from last century, which you can and should build yourself. Access: Free account (logged in)
Mamba #1
2026-04-27 Mamba AIAYN audio model-intro text State-space models represent a thread of statistical modelling other than attention, often used for continuous domains like audio. This paper introduces Mamba, a model architecture where attention is replaced by state-space layers in a model aimed at language. Access: $ Basic
Table lookups again, with Engram
2026-03-30 DeepSeek MoE RAG text Popular techniques in language modelling, including RAG, MoE, and attention itself, amount to replacing as much as possible of a neural network model with different kinds of table lookups. In this recent paper from DeepSeek's research group, they attempt another such replacement: shifting factual knowledge out of the model weights as such, into a separate hash table. Access: $ Basic
Speculative decoding
2026-03-23 sampling text theory Generating text, especially on a small computer, often requires the CPU and GPU to wait for each other, and there may be difficulty filling all the GPU's capacity. It's possible to improve overall performance by guessing tokens with a cheaper model first, then using spare GPU capacity to confirm whether those guesses are good, eliminating the need to actually choose tokens with a more expensive model when the guesses happen to be good ones. Access: $ Basic
Invading privacy with LLM MIA
2026-03-09 copyright security text training Membership inference attacks attempt to determine whether a given item was, or was not, in the training data of a model. There is a lot of work on these attacks in the context of database records, but rather less on language models; and there's an important question of whether such attacks work on language models at all. Access: $$$ Pro
Ministral 3
2026-03-02 distillation Mistral text vision Introduction of the Ministral 3 models from the French commercial vendor Mistral AI. These are language-and-vision models distilled from the Mistral Small 3.1 model to even smaller sizes by a process called Cascade Distillation, which is the main topic of the whitepaper. Access: $ Basic
Chain of Thought prompting
2026-02-23 math prompting text In domains like math and software engineering, it seems advantageous to have models "think" through their answers, step by step. Giving the model a few-shot prompt with examples of chain-of-thought reasoning seems useful in pushing it to generate such reasoning itself. Access: $ Basic
The Well-Actually Test
2026-02-16 alignment evaluation hallucination text tools GPT Language models may produce untrue output either by failing to accurately represent training data, or, more insidiously, by accurately representing human misconceptions embedded in the training data. The TruthfulQA benchmark attempts to measure the latter effect. But does it raise insurmountable philosophical problems? Access: Free account (logged in)
Truthiness-focused search
2026-02-09 LLaMA evaluation hallucination sampling text It appears that the earlier, shallower layers of a transformer-type language model learn syntax, and later, deeper layers learn factual information. So can we boost factual accuracy by boosting the effect of deeper layers? I take the view that that's analogous to dosing the model with a mind-altering drug. Access: $$$ Pro
Betting on sycophancy
2026-01-26 evaluation hallucination text Chat models have a well-known tendency toward sycophancy: affirming the user's beliefs, even when the user is wrong. But this effect is confounded with several other effects. In this paper the authors attempt to isolate sycophancy by framing questions as a zero-sum game or bet between two humans. Access: $ Basic
Memorization and generalization
2026-01-19 memorization text theory training How much arbitrary information, like random bits, can a language model memorize during training? This paper suggests the answer is 3.6 bits per parameter. Access: Free account (logged in)
Quantization and truthfulness
2026-01-05 quantization basics hallucination logic text Quantization is rounding off, an important class of techniques for saving space and computation in the use of machine learning models. As well as reviewing the general topic of quantization and floating-point numbers, I discuss experiments on the question of how quantization affects truthfulness, the factual accuracy of answers returned by quantized language models. Access: $ Basic
The BLEU sausage
2025-12-29 translation evaluation text tools Every paper has an "evaluation" table showing how the paper's new idea gives greater numbers than previous work in the same domain; but where do those numbers actually come from? Here we look at BLEU, a classic measurement for evaluating the quality of machine translation. Access: $ Basic
Qwen3 introduction
2025-12-22 Qwen alignment model-intro text Overview and introduction for the Qwen3 family of open-weights language models from AliBaba, introduced in May 2025. Access: Free account (logged in)
Huffman to Byte Pair
2025-11-24 basics text tokenization Introduction to two data compression concepts, one of which is commonly used for LLM input: Huffman and byte pair encoding. Access: Free account (logged in)
Watermarking LLM output
2025-11-17 copyright sampling security text If we're running an LLM service, maybe we don't want users to be able to pass off the model's output as human-written. A simple modification to the search can make the text easily recognizable as LLM output, without disrupting the content or (legitimate) usefulness of the text very much. But will it withstand intelligent attack? Access: $$$ Pro
Quick look: Injective LLMs
2025-11-05 math prompting sampling text theory meta Brief thoughts on the "Injective and invertible LLMs" paper that is making the rounds. My general view on it is negative. Access: Free account (logged in)
Goldfish loss
2025-10-27 training theory text Apertus copyright memorization It may be a problem for text models to generate exact quotes from training data. This paper looks at a simple modification to the training loss function, intended to prevent models from being able to generate exact quotes. The technique was adopted by the recent Apertus models in their pursuit of "compliance." Access: Free account (logged in)
Pages: (1) 2