Tim Tomov

PhD Student in Machine Learning @ DAML (TUM), how to know when models don’t know ? Doing uncertainty quantification stuff

[Tim Tomov @timtomov]

Are LLM outputs just language?

We often don’t know what the answer to a question will be - but we do know its structure. Numbers, lists, sets, ...

Our framework models outputs in their underlying structure for better answers + uncertainty.

w/ @dfuchsgruber @guennemann [1/4]🧵

[Andreas Kirsch 🇺🇦 @BlackHC]

This is an amazing read. Like AI 2027, it's a great way to spend some time thinking about the future and feels very plausible

That said, it's already overtaken by recent events thanks to the USG pulling Mythos and Fable already. This should help Europe wake up faster hopefully

[Tim Tomov @timtomov]

@SharonYixuanLi Interesting stuff ! In general I feel current methods do not really utilize the full information (distribution) the model has. Have you thought about Minimum Bayes Risk decoding baselines using the corresponding loss (Rogue, etc)?

[Probability and Statistics @probnstat]

Information Bottleneck Theory is a powerful framework introduced by Naftali Tishby for understanding learning and representation in intelligent systems. The central idea is that a good representation should compress the input while preserving information relevant to the prediction task. Formally, the objective balances two quantities: maximizing information about the target Y while minimizing unnecessary information about the input X.

A typical objective is:

I(T;Y) − βI(T;X)

where T is the learned representation and I(·;·) denotes mutual information.

In statistics and machine learning, the information bottleneck provides a principled approach to feature extraction, dimensionality reduction, and prediction. In deep learning, it offers a theoretical perspective on why hidden layers learn compressed representations that retain task-relevant information while discarding noise. It has influenced research on generalization, representation learning, self-supervised learning, and neural network interpretability.

In reinforcement learning, information bottleneck methods are used to learn compact state representations, improve exploration, and reduce sample complexity. The broader insight is that intelligence is not merely about storing information—it is about retaining the right information. By balancing compression and prediction, the information bottleneck provides a unifying principle connecting information theory, learning, and decision-making.

Image: https://t.co/jFnRTw87GD

[Adam Tauman Kalai @adamfungi]

1/4 Fixing hallucinations means fixing evaluations, as shown in our new paper https://t.co/Y75ldQUAbm building on our earlier @OpenAI blog. Accuracy-based scoring rewards models for making their best guess even when unsure, so hallucinations are like students guessing on tests.

[Tim Tomov @timtomov]

@mkirchhof_ It's interesting to explore the benefits and limits of defining an object's "form" vs. its information.

[Tim Tomov @timtomov]

@mkirchhof_ But this, of course, assumes you can find a map between a string and the distribution it represents, which is probably mostly not injective. But maybe another representation, like a credal set, could work. But in all these cases, you explicitly define the form of the object.

[Tim Tomov @timtomov]

So what?

While generally applicable, we think this could be especially useful for LLM-as-a-judge: better decisions ➡️ more reliable evaluations.

On a personal note: It was a really fun project, and I learned a lot!

[Tim Tomov @timtomov]

What if we could make decoding provably optimal?

Building on our recent work, we introduce Task-Aware Calibration, a framework for obtaining optimal LLM decisions for a given task.

The idea: Calibrate first, then act optimally

w/ @dfuchsgruber @raajeevVerma @guennemann

[Tim Tomov @timtomov]

So we get better decisions, but what actually changes?

It can, e.g., eliminate biases. In one LLM-as-a-judge task, the model often shows "correlation bias" (rating helpfulness and correctness identically). Task calibration recovers the true, weaker correlation structure.

[Richard Sutton @RichardSSutton]

The bitter lesson in 26 words:

Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.

[Jack Morris @jxmnop]

# A new type of information theory

this paper is not super well-known but has changed my opinion of how deep learning works more than almost anything else

it says that we should measure the amount of information available in some representation based on how *extractable* it is, given finite computation. for example, an encrypted text file has less V-information than the same data in plaintext, because it takes more computation to extract. note the contrast to traditional information theory, which would tell us that the two representations have the same amount of Shannon information

i’ve long wondered why certain types of basic questions didn’t have a proper theoretical answer:
> why does distillation outperform vanilla maximum likelihood training?
> why does lora work better than finetuning?
> why does self-attention work better than almost any other similar operation?
> how much “information” remains in a text embedding?
> how much “information” remains in language model weights?
> should i use fine-tuning or RAG?

the true answers to all these questions depend on some way of measuring and comparing *information content* between different representations. v-information is one step towards doing this

besides computational constraints, model architecture probably affects the “information content” in representations, along with the presence of any pretraining data used, as well as model-level statistics about the optimization in the training process – e.g. the length of time a model was trained for probably changes representations pretty drastically

this is all to say, i think there is some true notion of “information” that none of our current paradigms (Shannon information, V-information, etc.) capture. we encounter this idea every day but we dance around it and describe it in vague terms; we measure it from all sorts of angles but can’t quite characterize it theoretically

when the v-information paper came out I thought there would be a lot of follow up work developing more complex and useful notions of information for deep learning. but it hasn’t

yet I still think at the heart of these questions of what-information-lies-in-representations there’s something to be found that’s profound, elegant, and potentially extremely useful. I don’t know what it is and i’m probably not the person who will figure it out. but I really hope someone does. :)

[Tim Tomov @timtomov]

First paper of my PhD accepted @icmlconf in Seoul 🇰🇷 !

💡 Free-form language often has some underlying semantic structure. So why model LLM outputs still purely in text space?

Funnily enough: We leverage this view for our new paper, coming soon.

w/ @dfuchsgruber @guennemann

[Tim Tomov @timtomov]

Are LLM outputs just language?

We often don’t know what the answer to a question will be - but we do know its structure. Numbers, lists, sets, ...

Our framework models outputs in their underlying structure for better answers + uncertainty.

w/ @dfuchsgruber @guennemann [1/4]🧵

[Probability and Statistics @probnstat]

One theorem every ML engineer should know:

The Johnson–Lindenstrauss Lemma.

It states that high-dimensional data can be projected into a much lower-dimensional space while approximately preserving pairwise distances.

Why it matters:

• Explains why random projections work
• Enables scalable learning in high dimensions
• Used in embeddings, compressed learning, and ANN search
• Helps fight the curse of dimensionality

The surprising part:

You can reduce dimensions dramatically without destroying the geometry of the data.

That’s why many ML systems can operate efficiently even with massive feature spaces.

Modern representation learning is deeply connected to this idea:

Good embeddings preserve structure while compressing information.

In ML, compression is often not loss of intelligence —
it’s removal of redundancy.

[Eike Eberhard @ESEberhard]

New preprint:
"Transferable SCF-Acceleration through Solver-Aligned Initialization Learning (SAIL)"
* Standard ML initial guesses target the converged density.
* We target the solver trajectory instead.

[Lukas Gosch @lukgosch]

Very happy to present our work on a principled robustness evaluation of graph transformers at #ICLR2026! You can talk to us at tomorrow's poster session from 10:30 am to 1 pm in Pavilion 3 - #816. Joint work with Philipp Foth, @geisler_si, @SchwinnLeo, and @guennemann.