Dani Balcells

Machine learning interpretability researcher

[Dani Balcells @danibalcells]

@exploding_grad @zfjoshying Thanks for the breakdown!

[Dani Balcells @danibalcells]

@vintrotweets with @plasticlabs' take on LLM-native memory and how we're thinking of it as a reasoning task and the foundation of social skills for LLMs.

[Honcho @honchodotdev]

Second🫡Honcho Release Week Drop!

Memory as Reasoning.

Memory in agentic systems has historically focused on static storage, but we propose treating it as a dynamic reasoning task.

Humans evolved to leverage prediction & surprisal-based reasoning systems to deal with resource constraints.

LLMs and agents, however, don’t have these limitations, so we make the argument for logical reasoning as a trainable task to produce memory models that exceed human performance on several axes.

Scaffolding reasoning traces using this approach allows us to get more out of user and agent data and form more useful representations of personal identity.

This lays the foundation for something else really exciting coming later this week!

Read the full blog post linked below ⏬

[Dani Balcells @danibalcells]

Super cool work by @LeonGuertler and the TextArena team! Very excited to have collaborated on this :)

[León @LeonGuertler]

For the past ~2 months we have been working on training reasoning models on TextArena games. The first paper (introducing what we think is a very promising new paradigm) will hopefully be up later this week / early next; and the second one, focusing on the "scaling laws" of self-play and some additional analysis, tentatively around the 18th of july.

However, to get more feedback on the structure and implementation, we want to open-source the code now.

UnstableBaselines is a very simple Async, Online, Multi-Turn, Multi-Agent RL library built on vLLM and Ray. The code is pretty readable and around 1.2k lines long (and includes a cool rendering interface that you can run via "unstable-terminal")

1/7

[Dani Balcells @danibalcells]

Minimum vibable product.

[Dani Balcells @danibalcells]

@homsiT @readwise It would be my honor and craving.

[León @LeonGuertler]

TextArena is live on arXiv! We present a benchmark of 57+ competitive text-based games to evaluate and train LLMs on agentic behavior — including negotiation, deception, theory of mind and many more. Real-time TrueSkill. Multiplayer support. Human-vs-models. Model-vs-model. Perfect environment for Multi-Agent, multi-turn reasoning and Planning! [1/N]

[Dani Balcells @danibalcells]

@plasticlabs @variantfund @WhiteStarCap @betaworks @honchodotdev empire theory of mind 🌌

[Dani Balcells @danibalcells]

It's always possible that they would do a lot better with 100x more inference-time compute. Still, this is a different regime from coding or math, where the motivation for expensive reasoning models is that other models simply do a bad job. Here, Claude 3.7 is cheaper AND better.

[Dani Balcells @danibalcells]

However, it involves reasoning about cues that reasoning models might not be as sensitive to after SFT and RL: choice of words, personal style, mood... We wonder if the post-training that makes them so good at math and coding also makes them worse at reading social cues.