Martina Vilas

When to call it quits in LLM reasoning? 🛑

‪Martina's internship project suggests trace monitoring metrics and classifiers that can detect when an LLM reasoning trace is going to fail in mid way. The approach saves up to 70% of token usage, and it even helps with increasing accuracy by 2%-3%.

Working on this project was a great experience during my internship at @msftresearch.bsky.social 💙

Learned so much from this amazing team! Huge thanks to my coauthors: @vidhishab.bsky.social, Safoora Yousefi, @besmiranushi.bsky.social, @erichorvitz.bsky.social

We also found that these signals emerge EARLY in reasoning! At just 4k tokens, we can predict solution quality with ROC-AUC > 0.6.

This enables early path selection during parallel generation and ~60% token savings with +2.1% accuracy gains 🚀

Using LT signals for answer selection in multi-sample inference leads to:

⚡ 48% average token reduction (up to 70%!)
📈 +2.6% accuracy improvement over majority voting
🎯 Works by identifying correct paths even when the majority is wrong

Hidden states have distinctive temporal patterns for correct paths. They show:

✴️ Larger overall representational change (Net ↑)
✴️ Less wandering in latent space (Cumulative ↓)
✴️ More direct progress toward final state (Aligned ↑)

Across 3 reasoning models (DeepSeek-R1, Phi-4-Reasoning-Plus, Qwen3) and diverse domains (GPQA, AIME, TSP), LT signals:

✅ Significantly predict correctness
✅ Outperform output-based confidence measures and cross-layer signals

We track how representations evolve through the trace and extract 3 complementary signals:

📊 Net Change: Overall shift (start → end)
🔄 Cumulative Change: Total movement
🎯 Aligned Change: Progress toward final state

Identifying trace quality is critical: it enables more reliable predictions, improves efficiency by avoiding wasted compute, and can be used to guide models toward productive reasoning strategies.

Our solution: Look inside the temporal evolution of the model's latent space! 🔍

But not all reasoning traces are equal ⚖️ → some contain productive steps that lead to correct solutions ✅, while others deviate into overthinking, fail to converge, or exhibit inconsistent reasoning patterns ❌

Modern LLMs use chain-of-thought reasoning to solve complex problems, generating step-by-step solutions that can span thousands of tokens.

📈Scaling this inference-time compute (longer traces, multiple samples) significantly improves performance across reasoning tasks.

Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning Reasoning models improve their problem-solving ability through inference-time scaling, allocating more compute via longer token budgets. Identifying which reasoning traces are likely to succeed remain...

Can we predict which reasoning paths will succeed before seeing the answer? 🤔

Our new paper (arxiv.org/abs/2510.10494) proposes latent-trajectory signals from LLMs' hidden states to identify high-quality reasoning, cutting inference costs by up to 70% while maintaining accuracy

Eureka Inference-Time Scaling Insights: Where We Stand and What Lies Ahead - Microsoft Research Understanding and measuring the potential of inference-time scaling for reasoning. The new Eureka study tests nine state-of-the-art models on eight diverse reasoning tasks.

All Eureka inference-time scaling insights are now available here: www.microsoft.com/en-us/resear... It was fun sharing these and more together with Vidhisha Balachandran @vidhishab.bsky.social and Vibhav Vineet at #ICLR2025.

Looking forward to presenting this work next week at #ICLR2025! DM me if you are attending and want to grab a coffee to discuss these topics 💫

December 5th our ML theory group at Cohere For AI is hosting @mathildepapillon.bsky.social to discuss their recent review arxiv.org/abs/2407.09468 on geometric/topological/algebraic ML.

Join us online 💫

I’m putting together a starter pack for researchers working on human-centered AI evaluation. Reply or DM me if you’d like to be added, or if you have suggestions! Thank you!

(It looks NLP-centric at the moment, but that’s due to the current limits of my own knowledge 🙈)

go.bsky.app/G3w9LpE

I tried to find everyone who works in the area but I certainly missed some folks so please lmk...
go.bsky.app/BYkRryU

Does anyone know of any feeds (or similar) for student internship opportunities in ML/CV/NLP?

I've found starter packs on NLP, vision, graphics, etc. But personally, I would love to know and hear from researchers working on vision-language. So, let me know if you'd like to join this starter pack, would be happy to add!

go.bsky.app/TENRRBb

How do LLMs learn to reason from data? Are they ~retrieving the answers from parametric knowledge🦜? In our new preprint, we look at the pretraining data and find evidence against this:

Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢

🧵⬇️

LLMs tend to match problem-solving strategies based on textual similarity rather than truly understanding the underlying principles of mathematical problems.

Paper: Do Large Language Models Truly Grasp Mathematics? An Empirical Exploration From Cognitive Psychology

A starter pack of people working on interpretability / explainability of all kinds, using theoretical and/or empirical approaches.

Reply or DM if you want to be added, and help me reach others!

go.bsky.app/DZv6TSS

If you’re interested in mechanistic interpretability, I just found this starter pack and wanted to boost it (thanks for creating it @butanium.bsky.social !). Excited to have a mech interp community on bluesky 🎉

go.bsky.app/LisK3CP

👋 I also work on the field (examples on my profile). Would love to be added!

I forgot from whom in my feed I got this from, but anyway, this network analyzer is crazy efficient. It gives you ideas for accounts to follow based on your own followees. I just added 50 accounts or so.

bsky-follow-finder.theo.io

there are many smart speakers and thinkers around AI/ML and/or NLP. but i find almost everything to be kinda predictable by now, minor stylistic variations on the same story. who are some *interesting* speakers i should listen/read? i want things that may surprise or inspire me.

Any Latin Americans here working in Cognitive Science, very broadly construed? (Neuroscience, Psychology, Artificial Intelligence, Anthropology, Linguistics, Economics, Ethics, Philosophy, and more…)

I thought I’d create a starter pack but I could only find a handful of us. Say hi?

It is intuitive to observe some complex-looking model behavior (e.g., the classification of images of different animals using an abstract category) and infer an interesting capacity of the model (e.g., the ability to build rich representations that abstract away from particular animals).

We found that the mechanisms behind the emergence of these representations are similar to those of LLMs, and can be found across a variety of vision transformers and layer types.

Analyzing Vision Transformers for Image Classification in Class Embedding Space Despite the growing use of transformer models in computer vision, a mechanistic understanding of these networks is still needed. This work introduces a method to reverse-engineer Vision Transformers t...

[2/2] In our #NeurIPS2023 paper, we introduce a simple and efficient approach to investigate how class prototype representations emerge in vision transformers trained for image classification.

arxiv.org/abs/2310.18969

Position: An Inner Interpretability Framework for AI Inspired by Lessons from Cognitive Neuroscience Inner Interpretability is a promising emerging field tasked with uncovering the inner mechanisms of AI systems, though how to develop these mechanistic theories is still much debated. Moreover, recent...

We show how many of the issues in the AI Inner Interpretability field are similar to those in Cognitive Neuroscience.

We thus argue that we can adapt conceptual and methodological frameworks from CogNeuro to make progress in interpretability research.

arxiv.org/abs/2406.01352