Twitxy Saved

Polina Kirichenko

Verified account This account is verified. Learn more

@polkirichenko

AI researcher; prev. Research Scientist at FAIR @AIatMeta, visiting researcher at Princeton @VisualAILab, PhD at New York University
🇺🇦
San Francisco, USA polkirichenko.github.io Joined November 2018
1.4K Following 4.1K Followers 312 Posts

Latest posts 5

Polina Kirichenko Verified account @polkirichenko
Excited to release our work on *post-thinking*! We teach models to reflect on their generations and self-assess their own safety, correctness and confidence 🔎

Models can use their self-judgments to both select the best answers and learn to generate better ones!

🧵1/6
8 replies 15 reposts 132 likes 6.4K views View on X
Polina Kirichenko Verified account @polkirichenko
Check out our new work on overthinking in quantized reasoning models lead by superstar @LotfiSanae !
Sanae Lotfi Verified account @LotfiSanae Jun 25
When a quantized reasoning model fails, we usually assume it just lost the capability to solve the problem. But we surprisingly show that in up to 52% of failures, the quantized model actually reaches the correct answer but then does not commit to it. 🧵👇 1/8
View quoted post on X
0 replies 1 reposts 27 likes 4.4K views View on X
Polina Kirichenko Verified account @polkirichenko
Excited to release AbstentionBench -- our paper and benchmark on evaluating LLMs’ *abstention*: the skill of knowing when NOT to answer!

Key finding: reasoning LLMs struggle with unanswerable questions and hallucinate!

Details and links to paper & open source code below!
🧵1/9
11 replies 79 reposts 598 likes 134K views View on X