Cycles of compression-memorization during pretraining improves generalization:
https://arxiv.org/abs/2505.08727
Long context and long CoT are very different: the amount of information in long context makes it hard to combine all this inforation within the limited nb of layers, hence requiring long CoT, which are generated to "summarize" intermediate values during reasoning, thus enabling layers to use these intermediate values instead of the original ones.
https://arxiv.org/pdf/2505.04955
Most "maths" #LLM benchmarks only evaluate the final output (e.g., a simple number). There's a tendency though right now to shift maths evaluations towards assessing the reasoning path instead of the output. Both papers propose ways to do this, for theorem proving in Lean4, and judging the errors during reasoning:
ICL generalizes better than FT when usng the same training dataset, but if you further increase the
FT dataset with in-context inference, then FT ends up the best at generalization:
https://arxiv.org/pdf/2505.00661
Note: 2 artificial test sets are used, exploiting reversion of relation and syllogisms.
Transformers may encode a model of the World in their residual stream, that can be viewed as a space encoding the belief state (distribution over possible generating states): https://arxiv.org/pdf/2405.15943
Information compression in #LLM: the Kolmogorov Structure function is a lower bound of the scaling laws; assuming a Pitman-Yor data generation process (it's good to see Bayes again!), they demonstrate it leads to the actual scaling laws observed in LLM:
https://arxiv.org/pdf/2504.09597
ClusComp paper confirms that quantization seems to be hitting some ceiling, and proposes an alternative based on VQ codebooks:
https://arxiv.org/pdf/2503.13089
Overtrained #LLM are harder to finetune: this paper looks like a serious stop to
the current trend of small LLMs trained on more and more tokens.
And this was kinda expected: continued training of LLM without noise is difficult,
and there are previous papers about entropy of
activations that decrease during training, making learning more difficult, etc.
Optima from scaling laws means something after all. This phenomon needs further study, nice paper:
Edited #LLM knowledge may fail to propagate in multi-hop questions (e.g. in "the birth country of the father of...");
this paper analyzes the 2 circuits that are responsible for each reasoning "hop" and use them to edit knowledge:
https://arxiv.org/pdf/2503.16356
#LLM adaptation with LoRA is still too costly, and several works try to further reduce this cost.
To do that, LoRAM proposes to first prune the LLM and only finetune this prunes LLM with LoRA:
https://arxiv.org/pdf/2502.13533
This paper proposes a new metric to locate knowledge neurons in #LLM, based on
the change in target word logits when removing this neuron. It shows that both attention
and MLP encode knowledge, and that various types of knowledge is stored in different layers:
Job opportunity in Nancy, France: research engineer for pretraining LLMs. Candidate here: https://emploi.cnrs.fr/Offres/CDD/UMR7503-CHRCER-003/Default.aspx
Compression is prediction, and it is "closely related to the ability to generalize"; this is true for #LLM and Ilias Sutskever explains the effectiveness of unsupervised learning with compression. This paper exploits this connection to evaluate LLMs: https://arxiv.org/pdf/2402.00861
XLand-100B: a large scale datasets of 30,000 in-context reinforcement learning tasks trajectories
https://arxiv.org/pdf/2406.08973
https://arxiv.org/pdf/2501.09751 proposes an #LLM agent that progressively expands a knowledge tree, using search engines, and a conceptual pool which represents a kind of cognitive network, using reflection, in order to better answer a question. It can be thought as a very advanced RAG strategy.
This paper shows that #LLM in-context learning follows a scaling law (such as with training), and that structured representations of in-context samples emerge after a given scale: very reminiscent of training scaling laws isn't it? This also means that ICL may not work as well as it could if there's less than, say, 400 few-shot samples; hum...