PhD student @ The Hebrew University | Focused on tokenization, efficiency and multilingual LLMs
Joined August 2022
- Tweets37
- Following131
- Followers107
- Likes185
Pinned Tweet
"In a hole in the ground, there lived a hobbit."
BPE spends 13 tokens on this line, 7 of them on "in", "a", "the" and punctuation. CoBPE attaches those to their words: 6 tokens, same text.
Across English: 30% shorter sequences, better LLMs at the same compute. #COLM2026 🧵
Shorter sequences also mean real speedups.
Generating the same text takes about a third fewer decoding steps. Each step is a bit slower, but overall it's 1.28x faster.
Training on @karpathy's nanochat speedrun, CoBPE reaches GPT-2-level quality 1.23x faster.
Paper: arxiv.org/abs/2610.05597
Code and project page: co-bpe.github.io
Joint work with the wonderful @GKaplan38844 and @royschwartzNLP at @nlphuji.
Come to our oral on Thursday morning, or say hi at poster #126 in session 5, also on Thursday. #COLM2026
Yuval Reif @ COLM2026 retweeted
Can we tell when LLMs are being unfaithful in their chains of thought?
We evaluated 8 methods claiming to do this, and found that most perform near chance!
But evaluating this requires us to have ground-truth labels for CoT faithfulness. How can we obtain these?
Yuval Reif @ COLM2026 retweeted
Follow the Flow was accepted to #ACL2026 main! 🎉
T2I models often struggle with text-image alignment issues. These are usually addressed at the diffusion stage. We offer a different POV: analyzing and intervening at the textual encoding.
📄 arxiv.org/abs/2504.01137
Yuval Reif @ COLM2026 retweeted
Have LLMs become supervised learners (once again)?!
In our new paper, we argue that current LLMs’ post-training methods have effectively reverted to the "pre-train then fine-tune" era, explicitly tailoring models to desired behaviors.
1/n
🤖 Made with AI
Yuval Reif @ COLM2026 retweeted
Fine-Tuning LLMs on New Knowledge Encourages Hallucinations. (@zorikgekhman)
But why? We found something unexpected:
1M facts about city-like names →hallucinations explode.
1M facts about random identifiers →near zero!
Same model. Same number of facts. Only the names change.🧵
ALT "Why Fine-Tuning Encourages Hallucinations and How to Fix It" paper
Yuval Reif @ COLM2026 retweeted
Ever used a top-ranked LLM that just... felt wrong for you?
You’re not alone. Instead of leaderboards, many of us turn to "vibe-testing" - manually comparing models to our own needs. But can we turn these feelings into a structured evaluation?
New paper: "From Feelings to Metrics" 🧵
Yuval Reif @ COLM2026 retweeted
New paper: LLMs encode harmful content generation in a distinct, unified mechanism
Using weight pruning, we find that harmful generation depends on a tiny subset of the weights that are shared across harm types and separate from benign capabilities.
🧵
Yuval Reif @ COLM2026 retweeted
🚨New paper alert🚨
🧠
Instruction-tuned LLMs show amplified cognitive biases — but are these new behaviors, or pretraining ghosts resurfacing?
Excited to share our new paper, accepted to CoLM 2025🎉!
See thread below 👇
#BiasInAI #LLMs #MachineLearning #NLProc
Yuval Reif @ COLM2026 retweeted
🚨 New paper!
We present CHIMERA — a KB of 28K+ scientific idea recombinations 💡
It captures how researchers blend concepts or take inspiration across fields, enabling:
1. Meta-science
2. Training models to predict new combos
noy-sternlicht.github.io/CHI…
👇 Findings & data:
Yuval Reif @ COLM2026 retweeted
The longer reasoning LLM thinks - the more likely to be correct, right?
Apparently not.
Presenting our paper: “Don’t Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning”.
Link: arxiv.org/abs/2505.17813
1/n
Yuval Reif @ COLM2026 retweeted
Heading to @iclr_conf ✈️🧩 ‘Tokens→Words’ shows how LLMs build full‑word representations from sub‑word tokens and offers a tool for vocab expansion. 🚀
See our #ICLR2025 poster ‑ 26.4, 15:00‑17:30.
📄 arxiv.org/abs/2410.05864
🔗 guykap12.github.io/FromToken…
👇
📢Paper release📢 :
🔍 Ever wondered how LLMs understand words when all they see are tokens? 🧠
Our latest study uncovers how LLMs reconstruct full words from sub-word tokens, even when misspelled or previously unseen.
arxiv.org/pdf/2410.05864 (preprint)
👀 👇
[1/7]
Yuval Reif @ COLM2026 retweeted
[📄] Are LLMs mindless token-shifters, or do they build meaningful representations of language? We study how LLMs copy text in-context, and physically separate out two types of induction heads: token heads, which copy literal tokens, and concept heads, which copy word meanings.
Yuval Reif @ COLM2026 retweeted
✨ Ever tried generating an image from a prompt but ended up with unexpected outputs? Check out our new paper #FollowTheFlow - tackling T2I issues like bias, failed binding, and leakage from the textual encoding side! 💼🔍
arxiv.org/pdf/2504.01137
guykap12.github.io/guykap12.…
🧵[1/7]