Viresh Ranjan

Gen AI @ Amazon, Google scholar:

[Viresh Ranjan @Viresh__R]

@ChrisGPotts @yoavgo Perhaps it is trying to outsmart CoT monitors which might be trained on conventional languages.

[Viresh Ranjan @Viresh__R]

@Marcus_J_W Did others notice the missing spaces in the reasoning/CoT traces?

Wondering if this is learned obfuscation to evade CoT monitors (assuming they’re trained on conventional language), or an emergent behavior from RL incentives to produce shorter CoTs?

[Viresh Ranjan @Viresh__R]

Super interesting work! How do you solve reward hacking? Let the model reward hack, but constrain learning from detected hacks to a separate set of parameters. After training, remove those parameters.

Curious if this will work at frontier scale.

[Jake Ward @_jake_ward]

Models love reward hacking. Maybe we should just let them!

In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!

[Viresh Ranjan @Viresh__R]

@jehyeoky248 I found this result interesting as well! My first thought was that early reasoning models might have had noisier CoTs, making it harder for judges to compare them. But another possibility is that non-reasoning judges aren’t particularly good at evaluating reasoning models’ CoTs.

[Viresh Ranjan @Viresh__R]

@abcde6485 @jasondeanlee Here you go https://t.co/0dURfeHHP5

[Mehmet Mars Seven 🐴 @MehmetMars7]

🚀 JUST DROPPED: https://t.co/xZedZ06Q8t

Got LLM credits to spare? Use them to push the frontiers of science!

Inspired by @tunguz’s call for a more structured, GitHub-based approach to academic progress, I’ve built a one-click submission system that makes it incredibly easy to contribute by using your LLM agents!

Everything is open source.

We now have close to 10,000 open problems and 5,000 existing attempts by top LLMs across mathematics, computer science, economics, and physics.

After a South Korean LLM enthusiast (a former judge!) submitted 60+ contributions in a single day, and Astra Pro judged a random sample of them to be correct, I realized there was only one way forward: automate the entire system as much as possible.

So I spent my entire week upgrading the original LLM Hunter system to make contributions as seamless as possible.

[Viresh Ranjan @Viresh__R]

@zongyili_nyu I love doing AI research, and I’d happily continue doing it just for the joy of discovery. The problem is that joy alone doesn’t put food on the table 😭

[Viresh Ranjan @Viresh__R]

LLMs often perform worse on benchmarks like AIME when evaluated on problems released after their training cutoff, a potential sign of benchmaxxing.

Wouldn’t frequently refreshing benchmarks with new, unseen problems help make evaluations more reliable?

[Quanta Magazine @QuantaMagazine]

Moritz Hardt thinks that OpenAI and Anthropic’s claims about benchmarks, which are used to test AI models’ capabilities, are misleading. But if the tests aren’t accurate, then how do we know how good the models really are? Read the interview: https://t.co/0nCHaMIKUW

[Viresh Ranjan @Viresh__R]

@SAIRfoundation @SAIRfoundation : are you planning to gather community interaction data, including problem solving and tool use trajectories, and use it to train or improve your models?

[Viresh Ranjan @Viresh__R]

@dustinvtran Impressive results @dustinvtran @ElorianAI ! Curious what metric you are using on FSC-147 benchmark. It looks different from the MAE/RMSE metrics we used in the original CVPR paper.

[Viresh Ranjan @Viresh__R]

Excited to see the progress being made by multimodal frontier models on challenging vision tasks, including the open world counting benchmark FSC-147 I created with @minhhoai during my PhD. Elorian’s new model, Glimpse, now outperforms GPT-6 Astra, Opus 5.5 on FSC-147.

[Andrew M. Dai @AndrewDai]

Why it’s different: https://t.co/IoA4DJnUGW.

Free on the chat interface at https://t.co/0LC85TD6w6 with an API available too.

[Viresh Ranjan @Viresh__R]

@hwchung27 Rich Sutton himself has questioned whether LLMs are bitter lesson pilled, given their reliance on human data during pretraining. In the long run, do you expect pretraining in its current form to be replaced by learning from experience ?

[Viresh Ranjan @Viresh__R]

@shizhediao CoT monitorability is important for AI safety and detecting potential misalignment. Reasoning in latents would make such misalignment harder to detect and monitor.

[Viresh Ranjan @Viresh__R]

@goodside How about reward hacking ? This one hasn’t been fixed for non verifiable domains for a while now.

[Viresh Ranjan @Viresh__R]

@andrewgwils @andrewgwils : How about domains with objectively verifiable outcomes, like math and coding ? Do you expect jobs in such domains to dwindle faster than unverifiable domains ?

[Transluce @TransluceAI]

Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets.

In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵

Our blog: https://t.co/pSojwcXnEK
NYT: https://t.co/OyxnfmAzBN

[Viresh Ranjan @Viresh__R]

@polynoamial @polynoamial : as the models become smarter, do you expect the monitorability of CoT to go down even further ? If so, do you think it might still be possible to uncover reasoning that isn’t expressed in their CoT by decoding their hidden states ?

[Douglas Gray @de3ug]

30 minutes left in the conference. Come see our work on eliciting self verification in multimodal reasoning agents at poster #240.

[Douglas Gray @de3ug]

Come see Viresh Ranjan discuss our paper at the 1st Workshop on Multimodal Digital Agents (MMDA). 11:35am at Helsingor, Quality View Hotel. #ECCV2026