Nitay Calderon

327 posts

Nitay Calderon

Nitay Calderon

@NitCal

PhD candidate @TechnionLive | Google Research | NLP

Beigetreten Haziran 2022
770 Folgt437 Follower
Nitay Calderon retweetet
Gal Kesten Pomeranz
Gal Kesten Pomeranz@KestenGal·
Protein repeat detection is hard: repeated segments are often mutated and only approximately similar. Yet PLMs can still detect them well. But How? Check out our new preprint: "Induction Meets Biology: Mechanisms of Repeat Detection in Protein Language Models"
Gal Kesten Pomeranz tweet media
English
1
16
44
5.1K
Nitay Calderon retweetet
Zorik Gekhman
Zorik Gekhman@zorikgekhman·
New paper 🚨 We know that reasoning helps when step-by-step solutions are natural, for example in math, code, and multi-hop factual QA. But why should it help with factual recall, where no complex reasoning steps are needed? 1/🧵
Zorik Gekhman tweet media
English
3
16
90
13.4K
Nitay Calderon
Nitay Calderon@NitCal·
@AndrewLampinen Oh wow, definitely a very relevant and interesting work. Ill make sure to add a discussion of it in the newer version.
English
1
0
1
49
Andrew Lampinen
Andrew Lampinen@AndrewLampinen·
@NitCal Cool! You might like our work on latent learning: x.com/AndrewLampinen… — we similarly suggest that models fail to effectively use information when it needs to be recalled in a sufficiently different format, inspired by some LM findings but also other phenomena, e.g. in RL
Andrew Lampinen@AndrewLampinen

we argue that parametric learning methods are too tied to the explicit training task, and fail to effectively encode latent information relevant to possible future tasks, and we suggest that this explains a wide range of findings, from navigation to the reversal curse. 3/

English
1
0
14
1.3K
Nitay Calderon retweetet
roeeaharoni
roeeaharoni@roeeaharoni·
New work from our group on understanding factual errors in LLMs, led by @NitCal ! Models encode more knowledge than they can recall when prompted. We need to close this gap!
Nitay Calderon@NitCal

[1/7] Why do frontier LLMs make factual errors? Is it because they never learned the fact… or because they can’t access knowledge they already encoded? In our new paper, we show: The bottleneck is not encoding; it is recall. 🧵👇 Paper: arxiv.org/abs/2602.14080 Many thanks to @_galyo @bd_eyal @zorikgekhman @eran_ofek59358 @GoogleResearch

English
0
2
5
365
AK
AK@_akhaliq·
Google presents Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality paper: huggingface.co/papers/2602.14…
AK tweet media
English
5
13
52
11.3K
Nitay Calderon
Nitay Calderon@NitCal·
So, what can help recall? Thinking helps LLMs access the knowledge they already store. It recovers 40–65% of encoded-but-not-directly-known facts, with the largest gains for rare facts and reverse questions. This resembles the human “tip-of-the-tongue” effect. Bottom line: Given that encoding in frontier LLMs is nearing saturation while substantial headroom remains for recall, future improvements are likely to come not from scaling but from better utilization of existing knowledge. Paper: arxiv.org/abs/2602.14080
Nitay Calderon tweet media
English
0
1
4
217
Nitay Calderon
Nitay Calderon@NitCal·
[6/7] The reversal curse is when LLMs know "A is B" but can't answer "What is B?" If bidirectional knowledge were missing, reverse questions would be hard in any format. But in multiple-choice, reverse questions are as easy as direct ones. The problem, again, is recall.
Nitay Calderon tweet media
English
1
1
4
238
Nitay Calderon retweetet
Avi Caciularu
Avi Caciularu@clu_avi·
CoT is powerful but slow. @AmosaurusRex, my intern, introduces Thinking States—a way for LLMs to reason while processing input, not after. 🚀 Better latency 📈 (Almost) Matches CoT on complex QA 🧠 Learns from NL supervision Check out his thread 🧵
Ido Amos@AmosaurusRex

Can LLMs reason internally while processing their inputs, similar to how humans can think ahead as we process information? Our latest work introduces Thinking States, a novel architectural adaptation that transforms reasoning into a internal recurrent process. By training models to maintain a dynamic thinking state, we achieve significant inference speedups over Chain-of-Thought while substantially outperforming existing latent reasoning methods. Paper: arxiv.org/abs/2602.08332

English
0
7
27
4.8K

Entdecken

@AsaelSklar @GoldsteinYAriel @roireichart @CFGeek @_galyo @bd_eyal @zorikgekhman @eran_ofek59358