Manasi Sharma
@ManasiSharma_
Followers
344
Following
114
Media
13
Statuses
44
research engineer @scale_AI, working on reasoning for frontier models, agents, rl | prev @stanford, @StanfordAILab, @mitll, @Columbia
Stanford, CA
Joined July 2020
Happy CS Commencement Day! .S/o to Prof. @chrmanning for making CS 224N (Natural Language Processing) one of the most fun and rewarding courses that I have ever TA'd for😊.@stanfordnlp @StanfordEng @Stanford @StanfordAILab
2
5
61
going to ICML next week! let me know if you're interested in chatting about RL, reasoning, agents (esp. browser agents) and data + evals :).
3
2
98
new work on rlvr + natural language guidance to enhance a model's ability to reason in its own words.
What will the learning environments of the future look like that train artificial super intelligence? In recent work at @scale_AI , we show that training systems that combine verifiable rewards with multi-agent interaction accelerate learning.
1
1
6
first work since joining Scale is out!.
Adaptive Guidance Accelerates Reinforcement Learning of Reasoning Models. "Guide adaptively incorporates hints into the model’s context on problems for which all rollouts were initially incorrect and adjusts the importance sampling ratio for the "off-policy" trajectories in order
2
0
4
the Reasoning & Agents team at Scale AI is hiring researchers!🩵we publish research on new forms of supervision & training algorithms for reasoning models, agentic environments, & collaborate with top frontier labs to improve their models. feel free to DM me if interested!
3
0
2
excited to share that i have joined the frontier data research team at @scale_AI!🎉 i will be working on research in the domains of LLM reasoning, agents and rl.
0
0
6
RT @askalphaxiv: DeepSeek-R1 now offers an open source model with the ability to explain its intermediate reasoning. But how do explanation….
0
2
0
Amazing initiative by @rajpalleti314 and team!.
Goodreads for arXiv papers💡. What if instead of arbitrary algorithms and tweets, arXiv papers were curated by your research community?. Introducing communities on alphaXiv: bridging papers, discussions, and people in one space.
1
2
6
s/o @shaneguML for a very fun chat on multilinguality (and for taking the best pictures at the conference)
1
0
2
Had a great time at NeurIPS in Vancouver this year and met some amazing people in person!.
1
0
3
i’m going to be attending NeurIPS again this year! am super excited to meet new people and chat / learn more about: reasoning, post-training, alignment, agentic systems, code-gen (& please share any cool party invites!).
1
0
11
Been a while since I taught 224n, thank you for the kind words :).
Day 20 of NLP: .- finished cs224n lecture 19 .- it was purely python tutorial .- did a recap on numpy on:. > broadcasting . > matrix multiplication . > hardamard nproduct . @ManasiSharma_ is good instructor.
0
0
6
Open to any questions or comments!. Authors: Manasi Sharma, Ho Chit Siu, Rohan Paleja, Jaime D. Peña. 8/8.
0
0
1
Overall, we find that humans tend to trust differing model explanations equitably (and highly) in isolation, although post-hoc ones are trusted more when compared directly. This suggests important ramifications for human trust in plausibly misinformative model responses. 7/8.
1
0
1
We do not find a statistically significant tradeoff between explanation types and model performance, indicating that choosing explanation types that maximize user trust may not negatively affect model performance during text generation. 6/8
1
0
2
However, these gains disappear when users are shown responses independently, suggesting that humans trust all model responses fairly highly (~4/5 on a Likert scale), including deceptive ones, equitably when they are shown in isolation. 5/8
1
0
2
We find that any explanation from the model about its reasoning significantly increases (p ≤ 1.4×10-7) self-reported user trust when the user has the chance to compare responses. Trust is also higher when the justification follows the answer (post-hoc) in the retrieved case. 4/8
1
0
2