@paws_ed
Joined January 2018
  • Tweets684
  • Following391
  • Followers200
96Media
for years, I've been going through arXiv every day and will probably stop soon, there's 2k papers on cs.ml cs.cl and cs.ai today, i wonder if a paper or two from this will stand the test of time
2
55
Submission numbers at 52k 💀💀, even if 5-10k is covering for neurips results, thats like 2x last years ICLR submissions. Dont think 3 reviews per paper is going to happen, lucky if we reliably get 2
432
terrytao.wordpress.com/2026/…
39
Pawan Sasanka Ammanamanchi retweeted
Emily Riehl@emilyriehlSep 9
According to my memory from last night and the wayback machine, OpenAI updated the Navier-Stokes pdf at Tue, 08 Sep 2026 19:09:35 GMT. The original version was one page shorter and did not contain any citations to Diego Córdoba and Luis Martínez-Zoroa: web.archive.org/web/20260000…
13
134
14
1,314
63,377
Happy to have contributed to this work in a very small way, there’s a lot of cool analysis that can be done for eg: model ability based on the types of input, model failure despite instructions etc. We all think of model access equity but dont really consider language parity.
Vilém Zouhar@zouharviSep 4
Machine translation is not solved and it will take a while for it to be done arxiv.org/abs/2609.04173

Last Translation Benchmark

For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks...

arxiv.org
1
3
109
Its been a while since a translation benchmark addressed this gap beyond just wmt
31
Been using huggingface since the pytorch_transformer days, i personally think one of the big reasons pytorch was popular because of this library and the quick support for every major model, congrats to HF on the acquisition
47
Goemans' cost conjecture counterexample, Dinitz-Garg-Goemans, is actually a theorem. (Just a cursory Google search) Conjecture 1.2 mentioned in this paper arxiv.org/abs/2510.21287 (Goemans stated it long ago, and it first appeared in link.springer.com/article/10…)

Unsplittable Cost Flows from Unweighted Error-Bounded Variants

A famous conjecture of Goemans on single-source unsplittable flows states that one can turn any fractional flow into an unsplittable one of no higher cost, while increasing the load on any arc by...

arxiv.org
Dmitry Rybin@DmitryRybin1Jul 22
Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years. The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60. Chat with GPT 5.6 Pro where this was found: chatgpt.com/share/6a60b2eb-0…
1
185
posting here just for my curiosity about the problem's provenance
23
This benchmark probably costs a few thousand dollars to run on the absolute frontier.
Rosinality@rosinalityMay 28
arxiv.org/abs/2605.28079 Long context benchmark suite. It aggregates previous benchmarks.
2
97
Cool work, wonder what the cost of each solution is and if the answer is found in text first or lean first
Axiom
@axiommathaiMay 27
1/ Since February, 8 papers across algebraic geometry, representation theory, number theory, combinatorics have been quietly appearing on arXiv. Proofs by AxiomProver. 5 papers are now accepted at solid peer-reviewed math journals. To our knowledge, a first for the literature.
3
377
extremely impressive result, but it wouldve been very cool if oai also released the chain of thought of the model not just the final output to see all the model considered along the way. the discovery aspect is almost lost like this
OpenAI
@OpenAIMay 20
Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946. For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids. An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better. This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.
1
80
cdn.openai.com/pdf/1625eff6-… they do release it here
16
I am not sure how people are evaluating frontier models, or even open-source ones, these days. Running 35 problems on just a few settings [pass@1, pass@4, refinement] racked up the bill quickly to $100. Imagine trying to evaluate older NLP datasets with 10k problems+
2
71
Pawan Sasanka Ammanamanchi retweeted
Fabian Gloeckle@FabianGloeckleApr 3
A new milestone in automatic formalization: We translated an entire graduate math textbook into Lean using 30K LLM agents. Open-source, large-scale multi-agent inference that actually works > Blueprint+Lean: faabian.github.io/algebraic-… > Codebase+preprint: github.com/facebookresearch/… 1/7
23
137
14
704
91,844
Is evaluating an agentic harness the same as evaluating the base (post trained or not) model itself? Wonder if we’ll see a dichotomy. We care a lot about the ceiling capabilities today, not a lot of focus on improving/understand the floor.
38
Not to belittle the achievement, but how do you measure the alignment between what’s written in the proof vs what the agent wrote in the lean formalisation? What if the model went off on its own beat once it saw the formalised statement until the lean kernel accepts the proof?
Math, Inc.
@mathematics_incMar 2
We are pleased to share that using Gauss, we have completed a ~200K LOC formalization of Maryna Viazovska’s 2022 Fields Medal theorems on optimal sphere packing in dimensions 8 and 24. This is the only Fields Medal-winning result from this century to be completely formalized, and is the largest single-purpose Lean formalization in history. We are honored to have assisted @SidharthHarihar1 and the rest of the sphere packing team in this achievement. math.inc/sphere-packing

Completing the formal proof of higher-dimensional sphere packing — Math, Inc.

Dedicated to verified superintelligence via autoformalization.

math.inc
2
1
830
Deepseekv2 math is going to be great base model to train a theorem prover. But given that it already potentially saturates competition level math what will it be evaluated on?
82
Deep seek math v2 looks like a very exciting paper, the first of its kind when it comes to evaluating informal theorem proving. Training a verifier for proof generation and using it to train the model is something that's been tried in other domains as well.
1
1
314
more replies
They do later on use the automated verifier to generate annotation (especially for the final two loops of training), so bootstrap verifiers to make them good enough and then use them to annotate data? Paper: github.com/deepseek-ai/DeepS…

DeepSeek-Math-V2/DeepSeekMath_V2.pdf at main · deepseek-ai/DeepSeek-Math-V2

Contribute to deepseek-ai/DeepSeek-Math-V2 development by creating an account on GitHub.

github.com
1
1
113
We are definitely bottlenecked on hill climbing proofs because of needing human verifiers at the last stage.
37
Load more