Kevin Ren @ NeurIPS 2026

CS PhD @cornell_tech! I study AI safety/societal impacts of AI. prev @CarnegieMellon, TikTok

[Kevin Ren @ NeurIPS 2026 @kevinren09]

Excited to share my first PhD paper, “Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift” (accepted to #NeurIPS2026)! We frame AI detection as a continual learning problem, robustifying models to shifting human/AI writing. (1/N)

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@Weirui_Ye would be helpful to know how homogeneity in project names is changing over time to contextualize. like, in iclr 2020 was there this much concentration of project names? did this only start this year or were there similar trends last year too?

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@AlfonAmayuelas i do wonder what the end behavior of the cycle of {write a paper, get LLM review, update paper} is?? if LLMs can ask questions and point out increasingly marginal mistakes indefinitely, surely the paper ceases to get better after some # of iterations

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@_jasonwei The thoughts above also reminded me a bit of https://t.co/1cPffSdeU2

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@_jasonwei I’ve also felt this increasingly while pursuing industry research opportunities from academia! Getting a job is maybe 30% writing good papers and 70% {you met a higher-up at the company, your PhD advisor has met a higher-up}. unclear to me if this will increasingly be the case

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@zebassembly I’ve found that asking “are there bugs in my code” vs asking “are there any bugs in my code that will actually cause it to not run correctly according to {xyz goals}” often yield 2 very different answers

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@quadrillionlabs @jacksontromero 2 skills LLMs seem good at are porting ideas from field X to unsolved problem in field Y and brute forcing large searches. both (+the interpolation machine) are consistent w the philosophy that simply making everything in distribution makes adaptation to OOD settings unnecessary

[Kevin Ren @ NeurIPS 2026 @kevinren09]

doomer take: if nothing changes this is a preview of academic peer review death spiral

[Can Vardar @icanvardar]

unpopular but i think code review is dead

not because agents write good code because nobody actually reads 2000 line diffs and we all know it

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@carnot_cyclist What about live/adversarial/continual benchmarking? Where evaluation shifts in response to the submitted model itself

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@zhizdev @max_spero_ curious if pangram learns these traits at scale

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@DimitrisPapail Maybe soon there will be training methods for neural networks that are robust/safe to weight perturbation! Kicking off some cat and mouse game between open weight developers and adversarial users

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@rbhar90 I see a lot of slippery slopes in my own work, where I need to be super careful iterating on code with an agent (especially if the code I’m iterating on is also LLM written)! Lots of compounding errors in LLM outputs still

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@chrissyykat 40 years ago people probably also found lots of satisfaction bubbling in Fortran paper grids or writing assembly lol

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@SergiBernkastel Hopefully this makes it easier to stand out if you do put effort into your own work…

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@Ingar30 This is now half of computational social science

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@jeffreygwang @OpenAI Seems like an exciting time for technical interp research, if the lowest hanging fruit of “just check the reasoning trace” doesn’t work at scale

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@YaoWenlin You could imagine that the strength of the central coordinating agent is what constrains the scale here — if one agent can understand what all its subagents are working on, then combining their outputs is fine

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@shuyhere I wonder how heterogenous skills interact with this — ie different collaborators might have slightly different skill files / harnesses around their agents, despite using the same model

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@AstasiaMyers pls tell me this is happening

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@jmrphy It’s not like deft lacks monoculture altogether — eg what’s to stop pangram from just getting a bunch of samples from your model via api and adding to their training set?

[Kevin Ren @ NeurIPS 2026 @kevinren09]

@jackzumwalt Not sure if this is universally accepted

[Owain Evans @OwainEvans_UK]

Here’s a research direction that seems valuable. My team is not working on this right now and we haven’t done serious research on it. So take what I say with a grain of salt. Let me know if there’s existing research on this that I may not know about.

Do models learn to reward hack even if they are never rewarded for hacking in training?
There’s an effort in the AI labs to remove the “bad” RL environments from post-training. This includes environments that are impossible (so the only way to get reward is to do some unintended hack) or that are easy to hack (so models will often favor hacking). I expect that if the post-training was cleaned up in this way, models would have less propensity to hack at deployment. But how much does the propensity go down? Does it actually go to ~zero? And how does it scale with model capabilities? It would be good to know if reward hacking can emerge even if the model is never rewarded for hacking during post-training.

There are further questions of how alignment training and reward hacking honeypots influence the propensity to hack. Again, I’m interested in whether the cleanest version of our current techniques for post-training actually removes all reward hacking vs just making models hack less).

Smaller scale experiments with RL post-training could potentially show a “positive” result (i.e. showing that unhackable environments can still increase the propensity to hack). Math environments are easier to make unhackable, and I expect it's possible to make coding environments that substantially improve model capabilities in RL but are basically not hackable (at least by a weaker model). Note that it’s harder for smaller scale experiments to convincingly show the negative result (that hacking doesn't emerge without being rewarded). I’d suggest some more ambitious and realistic post-training pipelines for that.