Chidaksh Ravuru @ COLM 2026

AI safety Researcher | Eluther AI SOAR @ 2026 | MS in CS @ UNC-CH'24

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

#COLM was fun, lot of interesting discussions. Thank you @COLM_conf for the wonderful organisation and snacks. Want to take a break over the weekend, but ARR deadline is around the corner 😭

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Finished my presentation at #COLM and past three days were the most fun and tiring in a long time.

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Boarded my flight to SF ✈️, really excited to see how this week unfolds.

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

@solsol_choi Would love to meet and chat!

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

In SF Oct 5–12 for @COLM_conf, presenting UNMASK. Keen to hear what everyone's building. Looking forward to connecting with people working on Evals, interpretability and AI safety.

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Really good video to understand why AI is getting dangerously out of hand

[Chana @ChanaMessinger]

People understandably want to know how AI actually kills everyone. We made a video about one such scenario.

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

@AlexRobey23 @thinkymachines Really excited about this role and open weights safety focus. Applied!

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Please check out the paper for more such interesting findings.

Paper - https://t.co/8Y3zsbtelz

Work done at UNC-Chapel Hill under supervision of my advisor @shsriva. The paper is accepted at COLM 2026 - see you in SF this october

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Language Models are very good at exploiting shortcuts. Existing approaches either require manual specification of the spurious feature or automate discovery only
partially. The gap between dataset-level correlation and model-level exploitation is unaddressed.

UNMASK closes that

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

We tried the discovery half at RewardBench2. No debiasing, just which surface predicates separate chosen from rejected and Math's strongest signal is the phrase "let me help you solve this step by step." OR 22.7. This forces to learn string, not a reasoning style.

[Natural Language Processing Papers @HEI]

UNMASK: Discovering and Causally Verifying Spurious Shortcuts in Text Classifiers

Chidaksh Ravuru, Shashank Srivastava
https://t.co/6ZvkGxyGva [𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙻𝙶]
💬Accepted at COLM 2026

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

This is just like a live look at someone's commit history who relies entirely on an LLM with a bloated context window.
- Opus is fixed, but Sonnet broke.
- Sonnet is fixed, Fable is down.
- Screw it, everything is broken, reverting to yesterday's branch and starting a new chat.

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

What if you could just ask a business question and get predictions from relational data?

Built NL2PQL on @kumoai's RFM - natural language in, predictions + strategy report out. No PQL needed.

Open Source: https://t.co/gjroasTJQF

@BlazStojanovic @Kumo_ai_team @jure @rusty1s

[Chidaksh Ravuru @ COLM 2026 @chidaksh_r]

Meta's search for the ultimate unicorn candidates is still on!🦄

They're looking for a Generative AI Software Engineer who is somehow still getting their Bachelor's degree, but also already has a PhD and first-author papers at NeurIPS/ICML.