Chirag Raman

Assistant Professor @tudelft | Generative Multimodal ML, Embodied Social Intelligence, Digital Humans | Prev @LTIatCMU, @MSFTResearch, Disney Research, @cmuetc

[Chirag Raman @chiragraman]

@alexcliff15 Yup, if they were automatically flagged for two or more hallucinated references.

[Chirag Raman @chiragraman]

Finished my NeurIPS final metareviews and "use borderline scores sparingly" seems less of a guideline and more of a community-wide challenge mode.

Out of my batch of 10:
- 2 clear rejects
- 2 papers at 3.33
- 5 papers at 3.67
- 1 paper at 4 (with a reviewer confidence of 1)

My final justifications ended up averaging ~7-8k characters because the goal is anyway not to summarize scores but to assess based on the scientific discussion and, especially in case of rejection, provide authors with useful takeaways (as a PhD student, I always found superficial AC assessments more inexcusable than superficial reviews).

One thing that would make the process more productive is if reviewers were more explicit about what they think is required or remains unaddressed for a clear accept: which concerns are critical, what evidence would change their assessment, and remember to consider whether the core contribution is valuable to the community, even if there are remaining non-critical issues. Borderline scores are one thing, but borderline reasoning is much harder to synthesize.

[Chirag Raman @chiragraman]

@sanatan97 Just one; from a reminder to a reviewer who had said they would raise their score but hadn't officially. But I managed to get a few joint discussions with reviewers who clarified their position even though they retained their score.

[Chirag Raman @chiragraman]

@zussini

[Chirag Raman @chiragraman]

Given the discussion around the #NeurIPS2026 review process, some observations from my AC batch, looking only at papers where authors replied with sufficient time before the deadline (8/10):

• Only 3/8 papers had all reviewers return after rebuttal
• 5/16 participating reviewers first visibly engaged in the final 24h (3/16 in the final 12h) despite reminders
• Most 'discussions' ended after a single reviewer response

The discussion around 'long AI-slop' author rebuttals also feels more nuanced in my batch. Many long responses did involve substantial new scientific evidence, and needed the char budget to address the detailed questions.

Curious whether other ACs saw similar patterns.

[Chirag Raman @chiragraman]

@sanatan97 Yes this is usually the way to do it. I hope your AC factors in the content of the discussion beyond the scores

[Chirag Raman @chiragraman]

@sanatan97 Your reviewer(s) claimed too many new results in your response after asking for them?

[Chirag Raman @chiragraman]

Excited that SimPPL has been selected for @adaption_ai's inaugural research grant program!

I’ll be co-leading this with @swapneel_mehta, exploring adaptive representation techniques for investigative agents using Arbiter's real investigative traces as a grounded testbed.

The core question: how can agents move beyond treating documents, metadata, tool outputs, and analyst feedback as separate signals, and instead build adaptive representations they can reason over?

Grateful to @sarahookr, @sudip_r0y, and the Adaption team for supporting this work!

[Niloofar @niloofar_mire]

Tbh i’m kinda sick of this academic doomerism vibe consuming all of bay area and the self-aggrandizing pov that frontier labs have. Sure a lot of exciting stuff is happening but we wouldn’t be where we are wo academia & there is sth to be said about the pursuit of curiosity.

[Chirag Raman @chiragraman]

@shrimai_ @MistralAI @nvidia Congrats Shrimai! Looking forward to what you build with the team!

[Chirag Raman @chiragraman]

@Sahithya_Ravi @VeredShwartz @adityachinchure @LeonidSigal Cool work! We looked at something similar based on surprisal, maybe of interest:

Why Did This Model Forecast This Future? Information-Theoretic Saliency for Counterfactual Explanations of Probabilistic Regression Models https://t.co/gLXCtJXUE0

[Lucas Maes @lucasmaes_]

JEPA are finally easy to train end-to-end without any tricks!

Excited to introduce LeWorldModel: a stable, end-to-end JEPA that learns world models directly from pixels, no heuristics.

15M params, 1 GPU, and full planning <1 second.

📑: https://t.co/cpTzgvbTS0

[Haven Feng @HavenFeng]

✨Thinking with Blender~
Meet VIGA: a multimodal agent that autonomously codes 3D/4D blender scenes from any image, with no human, no training!
@berkeley_ai #LLMs #Blender #Agent 🧵1/6

[Google DeepMind @GoogleDeepMind]

We're helping AI to see the 3D world in motion as humans do. 🌐

Enter D4RT: a unified model that turns video into 4D representations faster than previous methods - enabling it to understand space and time. This is how it works 🧵

[Jonathan Stephens @jonstephens85]

What if you could turn a single 360° photo into a production-ready Isaac Sim environment in minutes?

That's exactly what we did here. Using @theworldlabs' Marble and an @insta360 X5 capture (rotating on top), we generated a complete navigable 3D environment and populated it with @LightwheelAI Sim Ready assets (bottom view).

The result? A fully interactive scene in Isaac Sim, ready for sim2real testing,. Navigation, manipulation, or any robotics task you need to validate.

What used to take weeks of manual 3D modeling and asset placement now takes minutes. Capture once in the real world, simulate everywhere in your training pipeline.

This is the future of robotics development with world models.

@NVIDIARobotics @nvidiaomniverse

#Sim2Real #Robotics #Simulation

[Chirag Raman @chiragraman]

@RitvikRastogi19 Congrats bro 🎉🎉

[Chirag Raman @chiragraman]

Attending the Embodied World Models workshop at #NeurIPS2025, DM if you're around and would like to chat!

Within the Dutch national consortium on Hybrid Intelligence, I co-lead the Robotic Surgery case study. This constitutes a multi-human-agent setting where a robotic agent participates in microsurgeries.

We have been incorporating bidirectional streaming and control between sensors in the environment and on-board the @NEURARobotics Maira, and @NVIDIARobotics Isaac sim, which serves as an imagined space for agents to reason and plan before acting in the real world.

Come say hi if this seems interesting or relevant!

[Chirag Raman @chiragraman]

@CoLLAs_Conf What's the best way to connect with Sarath, Razvan, or Eleni? Our NeurIPS paper MindForge (https://t.co/t7BxmclVi8) on perspective taking for lifelong learning sounds relevant, and my student who is a first author is Romanian, so would be cool to connect everyone.