Salt Lake City
Joined April 2014
- Tweets2.4K
- Following607
- Followers6.2K
- Likes9.1K
Ana Marasović retweeted
I feel quite sad from doing my ICLR reviews today. It appears that the grim reaper of AI has come for scientific publishing.
One paper was just desk rejected after I wrote a constructive review for it last week. The next two I had to give 1s. The fourth paper I recommended desk rejection, and the last I would reject for plagiarism, but the truth is that I don’t know if a human could attest to the writing.
Please, please try to address this problem, future PCs. The ship has sailed for ICLR 2027, but after this, we will need something to change.
Without a healthy peer review system, students are not exposed to unfiltered criticism of their work. This kind of criticism is essential for improving as a scientist, and it is a rising tide that lifts up the whole scientific community. We can’t abandon the system entirely. In order to be incentivized to engage with the system, reviewers either need payment or to be given work that is not complete drudgery to try to read. If reviewers disengage with the system or shirk their labor, as I very much want to do right now, this will lead to a death spiral where authors do not expect to receive fair reviews and reviewers do not expect to review genuine work.
I wish there were promising solutions to this problem that did not involve increasing the deadweight loss associated with paper submission, such as requiring videos with submissions or submission fees. But the cost of submission in real terms has gone down, and we will need to artificially raise it in order to even begin to cope with the surge in demand for AI publications. There’s also the “burn it all down” solution, but I find it hard to imagine what that would end in besides reinventing the current system.
Ana Marasović retweeted
Some thoughts on reviewing: spylab.ai/blog/reviewing/
First for @satml_conf where we trialed an early reject phase which cut submissions by ~40%.
Then more generally about how not everything deserves peer-review, and why we should reward writing things that people actually read.
Ana Marasović retweeted
KNOWS leaderboard is up! alexgill321.github.io/KNOWS-…
Ana Marasović retweeted
allowing papers like this is a terrible use of reviewers' time - it takes much longer to parse and is strongly correlated with the underlying paper being bad, which means we spend disproportionate amounts of time on bad papers
Ana Marasović retweeted
A few months ago, GPT-5.5 was the frontier model being run in OpenAI's web browser, Atlas. Now, budget models like GPT-6 Luna are outperforming Atlas at web use. The pace of AI progress is unreal!
KNOWS leaderboard is up! alexgill321.github.io/KNOWS-…
📄 Paper: arxiv.org/abs/2609.30604
🌐 Project page: alexgill321.github.io/KNOWS-…
💻 Code: github.com/alexgill321/KNOWS…
Ana Marasović retweeted
Okay, so someone who was "some PhD student" who made a viral benchmark and someone who worked on product evals at DeepMind for Gemini, I think I might have a unique perspective here.
🧵
META’S HEAD OF AI SLAMS AI BENCHMARKS
"it’s some PhD student who made these benchmarks"
"you’re kind of outsourcing all of your product thinking to this viral benchmark"
"so clearly that’s wrong"
I think there is something to this. the whole industry developed so rapidly that there was hardly any time to establish proper benchmarking. companies used what was available, competitors then had to catch up on the same benchmarks, and some of them eventually became industry standards
apparently there is quite a lot of randomness and chance in how we are measuring AI progress today
I too am at #COLM2026! (first conf since july 25, oops)
I'm not presenting anything, but we just released a benchmark (👇), and I'd love to chat anything AI safety (CoT monitoring esp), interpretability, simulating users, and agents
FWIW this is the most ambitious benchmark I've co-created!
𝗞𝗡𝗢𝗪𝗦 jointly & reliably evaluates agents on:
1) live web search,
2) use of common web productivity tools,
3) visual/spatial understanding;
programmatically where possible and by LLM judges where needed, more 👇
Ana Marasović retweeted
Three papers from our lab at #NeurIPS2026, spanning LLM reasoning faithfulness, belief formation, and interpretability of MoEs:
⭐ Faithfulness Metrics Don't Measure Faithfulness: A Meta-Evaluation with Ground Truth
arxiv.org/abs/2605.25052
We introduce a benchmark for CoT faithfulness metrics, and find that existing metrics are either close to random or prohibitively expensive to run
@GurYoav @anmarasovic
⭐ Indications of Belief-Guided Agency and Meta-Cognitive Monitoring in Large Language Models
arxiv.org/abs/2602.02467
An interdisciplinary collaboration applying the HOT-3 indicator of consciousness to LLMs using interpretability tools, revealing surprising patterns in belief formation and action selection in LLMs
@noam_steinmetz @GoldsteinYAriel @Liad_Mudrik
⭐ Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts
arxiv.org/abs/2605.12476
We find a geometric coupling between routers and their corresponding experts in MoEs, explaining how routers form assignments that support an effective division of labor
@sageaarc @noyahochwald
Ana Marasović retweeted
My hot take is that if your company builds software that it cannot stop from hacking into other systems, your company is a malware company and should be treated as such
Ana Marasović retweeted
I found this argument extremely thoughtful: the only CoT that will survive is the CoT models need to solve hard problems, so we should limit latent reasoning architectures.
The piece expresses the view that interp will not save the day "in the near future". That is reasonable, but I feel strongly that interp is something you'd want to nurture as a hedge against the possibility that people (or AIs) will develop latent reasoning architectures anyway. Their Appendix C says to me that interp should be more focused on latent reasoning models, especially if token-level monitoring suffices for other models.
If (when) CoT becomes unmonitorable, we have either developed other defenses or not. More generally, we should have contingency plans even for unlikely and/or hopeless-seeming outcomes.
We've written a post arguing that latent reasoning architectures (aka 'neuralese') would substantially increase misalignment risk via making oversight much harder.
In the extreme, we could see massive 'neuralese hivemind swarms' where the agents think and communicate in latents, likely making oversight nearly entirely reliant on observing the actions these agents take. (And these agents would have huge amounts of time to reason about obfuscating their actions if they wanted to do so...) Individual agents doing extensive latent reasoning would also be concerning; in the post we discuss how above some threshold of latent reasoning, agents may be able to perform difficult-to-detect and reliable steganography for communication and further reasoning.
We argue both that latent reasoning architectures would make chain-of-thought no longer very useful for oversight (by eliminating or greatly reducing the need for verbalized reasoning) and that, without these architectures, it's likely the value of chain-of-thought for oversight could be preserved.
redwoodresearch.org/blog/lat…
I hear people are already getting a new type of spam daily from random people with "single-authored" papers asking them to be their ARR service contributor 💀💀💀
Ana Marasović retweeted
😢 The marginal cost to generate and submit a paper is essentially zero, but the human cost to evaluate it remains brutally high.
I seriously think that we're going to need to bring back viva voce just as an anti-spam CAPTCHA.
Peer review is a gift economy built on the assumption that everyone is acting in good faith, but it's not going to be able to survive when that faith is assymetric.
Yet another evidence for this opinion:
If your venue needs 3 human reviewers to read slop before it gets desk rejected, and you still expect them to keep reviewing for you, you're doing it very, very wrong.
Similarly, if you allow slop reviews, quality work will go elsewhere.
TMLR has faced a deluge of submissions, necessitating stricter desk rejection policies due to limited reviewer capacity
Co-EiC Nihar Shah reached out to authors of 10 papers slated for desk reject. Could they answer questions about their *own* submission?
medium.com/@TmlrOrg/asking-a…
