@biunlp

The Bar-Ilan University, Natural Language Processing group.

biu-nlp.github.io/
Joined August 2011
  • Tweets240
  • Following108
  • Followers800
1Media
BIU NLP retweeted
Mosh Levy@mosh_levyOct 8
Now published in Nature Machine Intelligence! nature.com/articles/s42256-0… Huge thanks to the reviewers for a constructive review process. The paper is better for it.

Characterizing the role of reasoning tokens in large language models as state over tokens

Nature Machine Intelligence - When large language models ‘reason’ about a problem, they might still solve it differently from what their reasoning transcript suggest. In this...

nature.com
Mosh Levy@mosh_levy16 Dec 2025
New paper: We are often told that reasoning tokens aren't faithful explanations. But to have a useful metaphor for their operation we need a characterization of what they are, not what they are not. To that end, we suggest "State over Tokens" (SoT) 👇🧵
2
3
28
590
BIU NLP retweeted
omer goldman@omerNLPOct 6
We trained an LLM for Yiddish! Existing pretraining data is only about 40% real Yiddish and models miss a lot of what makes the language what it is. MameLoshnLM fixes quite a bit of that! Come hear @UrikaUri talk about it at #COLM2026. Tmrw 11am Grand Ballroom Poster #123
Uri Katz @ COLM 🌉@UrikaUriOct 6
Replying to @UrikaUri
Yiddish combines Germanic grammar, Hebrew script, and fusion vocabulary. Multilingual LMs can produce readable Yiddish while missing its distinctive vocabulary and grammar. The result is a flatter version of the language.
3
3
2
31
2,508
“A language is a dialect with an army...,” goes the saying about Yiddish. For LMs, training data may be that army. When native texts are overshadowed by noisy translations, multilingual LMs can look fluent while missing what makes a language unique. MameLoshnLM. #COLM2026 🧵
3
4
17
1,265
In deep research reports, many sentences aren't supported by their citations: recall is 58.7% in NVIDIA AI-Q and 7.1% in TrajectoryKit. Our #EMNLP2026 paper (main conf) proposes an algorithm that traces each error to the agent that introduced it, and identifies the error type.
3
15
1
35
2,682
Paper: arxiv.org/abs/2608.24306 @huggingface: huggingface.co/papers/2608.2… Code: github.com/eranhirs/who-is-t… Collab w/ @meetdavidwan @HanWang98 @EliasEskin @mohitban47 Ido Dagan @biunlp @unc_ai_group See you in Budapest! 🇭🇺

Who is the Agent to Blame? Localizing Faithfulness and Citation...

Deep research (DR) systems produce long-form cited reports by orchestrating multiple agents that search and synthesize information from the web. Citations are the primary mechanism for evaluating...

arxiv.org
1
8
202
BIU NLP retweeted
Yanai Elazar@yanaielaSep 8
I'm looking for an intern to join my group, around February next year for four months, through the Azrieli PhD Fellowship. DM/email me if you're interested! Link in the next tweet.
5
33
230
19,375
In conversations we often refer to things without naming them directly: "the change that broke token refresh" or "the issue Amy opened yesterday" and expect others to know what we mean. Can agents infer what we're referring to and actually find the right item using its tools?
1
7
2
26
6,539
Mark your calendars for ISCOL '26! 🤩 We're excited to announce that this year's ISCOL conference will take place on December 23rd at The Hebrew University. See you soon in Jerusalem:)
1
11
1
28
1,638
For more details: iscol-meeting.github.io/isco…

ISCOL 2026 — The Israel Seminar on Computational Linguistics

Join ISCOL 2026 on December 23, 2026 at Hebrew University of Jerusalem for exchanging ideas in Computational Linguistics and NLP across academia and industry in Israel.

iscol-meeting.github.io
2
4
188
🚨 We present 𝐌𝐀𝐍𝐂𝐄: 𝐌𝐀𝐍𝘪𝘧𝘰𝘭𝘥 𝘢𝘸𝘢𝘳𝘦 𝐂𝘰𝘯𝘤𝘦𝘱𝘵 𝐄𝘳𝘢𝘴𝘶𝘳𝘦 (with @yoavgo and @yanaiela). MANCE is a new state-of-the-art method for concept erasure.
2
20
1
57
9,841
🚨 Presenting PrefixNLI (Oral) tomorrow at #ACL2026! 🌴 Hallucinations start with a single token. So why wait for a full sentence to detect them? We extend Natural Language Inference to arbitrary text prefixes, detecting factual inconsistencies as soon as they arise and using that signal to guide decoding. Our controlled decoding method delivers major faithfulness gains in summarization. A 3B model matches an 8B model in faithfulness while maintaining similar runtime, using half the memory, and is substantially faster than prior lookahead methods. @SapirHarary, @hirscheran, @lovodkin93, @meetdavidwan, @mohitban47, Ido Dagan 📍 Regatta ⏰ 2:10pm arxiv.org/abs/2511.01359 Hope to see you there! 👋 #NLProc #ACL2026

PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise

Natural Language Inference (NLI) models have been used in various ways to improve the factuality of LLM outputs. This is typically done by applying an NLI model to judge whether the model output...

arxiv.org
1
7
1
24
1,875
Replying to @mohitban47 @aclmeeting @danadaeun @jwlee8877 @duynguyen772 @niu_tianyi @EliasEskin @UTCompSci @icmlconf @cyjustinchen @hyunji_amy_lee @hanqi_xiao @jmin__cho @jhuclsp @jaeh0ng_yoon @NTUsg @ArchikiPrasad @codezakh @gentaiscool @zhan1624
-- PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise (Oral) @SapirHarary , @hirscheran , @lovodkin93 , @meetdavidwan, Ido Dagan @biunlp arxiv.org/abs/2511.01359 #ACL2026 Regatta Session 3: Oral Session B: Sun. July 5, 14:00-15:30 PT xitter.cf/SapirHarary/status/198…
Sapir Harary Israel@SapirHarary7 Nov 2025
🚨 New paper alert! We’re thrilled to share our new preprint “PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise” ✨ LLMs generate text one token at a time, but factuality checks still wait for a full sentence. We extend NLI to text prefixes, enabling the catching of factual errors as soon as they arise during generation, achieving major faithfulness gains with 25× lower computational cost than lookahead methods. 🧵⬇️
1
5
15
726
BIU NLP retweeted
Mosh Levy@mosh_levyJun 11
New paper! People treat reasoning trajectories as text, but what if we can do better than that? We show that we can, by training Behavior Forecasters (BFs) that get a reasoning trajectory as input and make more accurate forecasts than frontier models at a fraction of the cost. 🧵
3
22
2
174
14,386
BIU NLP retweeted
Yanai Elazar@yanaielaApr 14
Are you interested in interning with me and my lab? A unique opportunity for a 4-month research stay, with generous funding as an Azrieli visiting PhD fellow! DM me if you're interested. azrielifoundation.org/fellow…

International Visiting PhD Fellowship - The Azrieli Foundation

The Azrieli Fellows Program is an elite group of academics who cultivate a network of leading professionals in Israel and around the world.

azrielifoundation.org
9
49
2
270
29,081
BIU NLP retweeted
Yanai Elazar@yanaielaJan 29
🚨 New Study 🚨 @arxiv has recently decided to prohibit any 'position' paper from being submitted to its CS servers. Why? Because of the "AI slop", and allegedly higher ratios of LLM-generated content in review papers, compared to non-review papers.
4
20
5
104
26,610
If we want to understand LLMs' reasoning, we must stop reading tokens as text and start decoding them as state to reveal how models actually use them. In colab with: @ElyosephZohar, @ravfogel and @yoavgo. Paper: arxiv.org/abs/2512.12777 Blogpost version: mosh0110.github.io/blog/stat…

State over Tokens: Characterizing the Role of Reasoning Tokens

Large Language Models (LLMs) can generate reasoning tokens before their final answer to boost performance on complex tasks. While these sequences seem like human thought processes, empirical...

arxiv.org
4
6
4
41
6,713
New paper: We are often told that reasoning tokens aren't faithful explanations. But to have a useful metaphor for their operation we need a characterization of what they are, not what they are not. To that end, we suggest "State over Tokens" (SoT) 👇🧵
5
18
5
65
14,699
Registration for ISCOL is now open! Please register as soon as possible, as spots are limited and we need to get an accurate headcount. iscol-meeting.github.io/isco…

ISCOL 2025 — The Israel Seminar on Computational Linguistics

Join ISCOL 2025 on December 18th at Bar-Ilan University for exchanging ideas in Computational Linguistics and NLP across academia and industry in Israel.

iscol-meeting.github.io
3
2
6
2,315
PrefixNLI introduces fine-grained factuality detection through prefix-level entailment, laying the groundwork for extensions across tasks, domains, and training paradigms. Collab w/ @hirscheran , @lovodkin93 , @meetdavidwan , @mohitban47 , and Ido Dagan @biunlp @unccs @unc_ai_group 📘 Paper: arxiv.org/abs/2511.01359 💻 Code, datasets, and models: github.com/sapirharary/Prefi…

PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise

Natural Language Inference (NLI) models have been used in various ways to improve the factuality of LLM outputs. This is typically done by applying an NLI model to judge whether the model output...

arxiv.org
3
12
350
🚨 New paper alert! We’re thrilled to share our new preprint “PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise” ✨ LLMs generate text one token at a time, but factuality checks still wait for a full sentence. We extend NLI to text prefixes, enabling the catching of factual errors as soon as they arise during generation, achieving major faithfulness gains with 25× lower computational cost than lookahead methods. 🧵⬇️
4
18
4
63
8,304
Load more