Alexander Chemeris @ ICML

The gap between Time-Series analysis and the rest of AI is so painful that I decided to research in public, publishing experiments, thoughts, and interesting papers to stir things up a bit. Comment or follow if you believe the future is just the next step in time series.

I love flying — it gives me enough distraction-free focused time to learn something new. I've been wanting to learn how flow matching works for modelling complex probability distributions in data, as it's critical for modelling telemetry and infrastructure failure modes. On my flight from ICML back to Cape Town, I listened to the flow matching lectures by @peholderrieth. It's the best explanation I've seen — very well structured. Highly recommend: diffusion.csail.mit.edu

Strongest impression from ICML: main-track talks often feel like award ceremonies. The work is ~1 year old; the author has moved on. Workshops feel alive because the lag is shorter. You’re talking to people still wrestling with the problem, not just presenting finished work.

@MetatrolN @fin_presented Awesome work!

AI4Math research focuses on Lean autoformalization and theorem proving. But how do we come up with new conjectures? In our work with @fin_presented, we study how well can LLM agents use computation to create hypotheses and solve research-level mathematical problems? arXiv: arxiv.org/abs/2607.06820 We evaluate 15 LLMs in zero-shot and SageMath-augmented agentic setup on RealMath 133 problems extracted from arXiv mathematical papers. Key findings: - Tool access improves every model, by 9.7 pp on average. Open-weight models gain 15.3 pp, compared with 6.5 pp for closed frontier models, narrowing the gap between them. - Most intriguingly, a CAS-augmented agent reproduced a computational mathematician’s workflow: computing intermediate objects, finding patterns, forming conjectures, recovering from errors, and validating formulas across parameters (see the details in the case study). - Gains vary: #Qwen 3.7-Max rises from 42.1% to 69.9%, a gain of 27.8 pp that brings it close to frontier performance. #Kimi 2.7 gains only 1.5 pp. - Tool-use behavior is strongly bimodal. Strong agents usually finish in 3-4 tool turns, while weaker agents often exhaust all tool budgets. - The largest gains are in combinatorics (+18.7 pp) and rings and algebras (+10.7 pp), while algebraic topology and group theory remain difficult. - Recovery after a failed tool call ranges from 16% (#Sonnet-5) to 77% (GPT-5.5) across models. The ability to revise a strategy after receiving computational feedback separates effective agents more clearly than the raw number of errors. Interesting observations about some models: - #GPT 5.5 leads in both solve rate and efficiency, reaching a 75.2% accuracy with the lowest token usage among tool-enabled agents. - #MiniMax M3 is the least efficient, using the most tokens per problem and achieving substantially lower accuracy. - #Opus 4.8 exceeds Opus 4.7 by only one solved problem. - #Grok 4.3 shows one of the worst results and produces 248/336 SyntaxErrors🥲. - #Fugu Ultra shows the smallest increase in token usage with tool access, at 4.5×, averaging 70k tokens per problem. Today we are presenting our poster at the @ai4mathworkshop at @icmlconf. Come by to discuss our work. #ICML2026 #AI4Math #AI #Agentic #LLM #Mathematics

@MParakhin @dylan522p If you're still at ICML, I'd be very curious to hear a long answer 😅 I'm thinking about this a lot, as what I'm doing at langotime.ai depends on the answer a lot...

@chemeris @dylan522p I have given it a lot of thought, not easy to condense into a tweet, though.

Exactly 2 years ago, at ICML in Vienna @dylan522p and I made a bet on who will have more training compute deployed in two years (now): OAI/MSFT or Anthropic/AWS. I claim I won that bet (OAI), Dylan! :-) Care to have another one? Ready to take you up on below :-)

Actually, I like this version more, but it's not on brand..

A non-technical friend has asked me to explain what langotime.ai is about in pictures. The first iteration with Claude prompting gpt-image-2. Not great, not terrible, but has promise, I guess 😅

Two of my papers were accepted to @kdd_news MILETS workshop! Blog posts coming soon. 1. LeNEPA: No-Augmentation Next-Latent Prediction for Time-Series Representation Learning (arxiv.org/abs/2607.00958) 2. Aionoscope: Debugging Latent-State Accessibility in Time-Series Representations (arxiv.org/abs/2607.00956)

@xia_char @denise_teng25 @shawnshenjx @FusionFundVC @GradientVC Hey, I'd love to join. I'm founder @ langotime.ai, building Robots for your Digital Infrastructure like LLM inference and K8s clusters. Sent a request @ Luma

Hosting a happy hour🍸during #ICML2026 in Seoul with @denise_teng25 and @shawnshenjx. Join us for an evening of conversation with founders and researchers pushing past scaling laws. 📅July 9 | 7pm | Oakwood Hotel (COEX) — details on RSVP @FusionFundVC + @GradientVC + @memoriesai RSVP link in comment to secure your spot!

@richard_qiuyi_z @ElorianAI Hey, I'm building multi-modal models to reason about telemetry (not vision) in real-time @ langotime.ai, but challenges are very similar. I've sent a request @ Partiful, but would be happy to catch you at the conference if the event is full

Coming to ICML and interested in the next gen multimodal models? Come hang out with @ElorianAI at ICML!

@nikhil_suresh2 @ElorianAI @RicursiveAI @DeltaInstitutes @strikervp Hey, I'd love to join. I'm founder @ langotime.ai, building Robots for your Digital Infrastructure like LLM inference and K8s clusters. Sent a request @ Partiful

@ElorianAI @RicursiveAI @DeltaInstitutes @strikervp link below: partiful.com/e/7qj7GW6CtIaX…

ICML week in Seoul!! 🇰🇷🇰🇷 We are hosting a rooftop dinner with @ElorianAI and @RicursiveAI. Amazing food, cool views, and great people. Come hang out! Thanks @DeltaInstitutes for co-hosting with @strikervp.

@DeltaInstitutes @RicursiveAI @ElorianAI @strikervp Hey, I'd love to join. I'm founder @ langotime.ai, building Robots for your Digital Infrastructure like LLM inference and K8s clusters. Sent a request @ Partiful

🇰🇷 Headed to ICML 2026? Come hang out with our friends at @RicursiveAI, @ElorianAI, and @StrikerVP to celebrate the start of the main conference!!

@matrosov Better than Opus 4.8 in your opinion? And how does it compare to GPT-5.5? I'm using Opus to plan and GPT to execute r/n since it's much more robust

Sonnet 5 is a huge leap forward in coding, agentic tasks, and adaptive thinking. Give it a try.

If you're not a GPU-rich lab, do time-series research. Smaller models than vision. Faster iteration loops. Real applications right away instead of theoretical demos. I find that if an experiment takes >1h to run, I'm not moving fast enough.

@MParakhin Well, a year from now both will look like baselines 🤷 I guess we just need to train patience

Not as relevant now :-(: I had an opportunity to deeply test both Fable 5 and GPT-5.6 Max. 5.6 is clearly better than Opus 4.8 at everything (slightly faster, too, though that depends on the load). Vis-a-vie Fable, it is clearly worse on coding, but better on agentic workloads. I had Fable write code, 5.6 run experiments - dreamy…

Awesome work on JEPA for real-time control

@DimitrisPapail I guess if benchmarks reported individual skill mastery values rather than a single blended value, we would see a lot more interesting picture.

The path from VibeBlogging to Science is a couple of months long :D microsoft.github.io/benchpress/add…

Awesome work! Have you thought about training it to choose the width at inference time - to maximize quality at a given FLOPs budget? So that the shape is the result of the model's training and could be different for different types of the input. Recent papers from @sainingxie team showed that e.g. text and images "propagate" differently through a transformer, so the most optimal shape is likely different. I further suspect that even within a a single modality inputs of different complexity "propagate" differently. I'm tracking this idea here: alex-wiki.langotime.ai/ideas/hierarch…

Also, I suspect that X-shape is because of the token-space reconstruction training. I suspect that for JEPA-like training, it'll likely look /\-shaped. At least that's what I see from the representation structure research papers, and from looking into the (time-series) models representation structures across layers: blog.langotime.ai/about-manifolds aionoscope.langotime.ai/manifolds aionoscope.langotime.ai Have you looked in to this?

Introducing ><former Most transformers are rectangles◻️: every layer has the same width But is that optimal?🤔 We propose variable-width transformers that have different widths across layers, improving loss while cutting compute & KV cache size 🧵