Vaishnavh Nagarajan

Congratulations to CSD faculty Aditi Raghunathan and her research collaborators on receiving an ICML Outstanding Paper award for Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction (icml.cc/virtual/2025...).

Paper: arxiv.org/abs/2504.15266

Reading the dedications of a PhD thesis is often a cure for a bad day. There’s so much affection in them

As NeurIPS review deadline is around the corner, please remember that you cannot use any non-local LLM like chatgpt/gemini for understanding the paper and drafting/revising your review as that breaks the confidentiality agreement.

NeurIPS 2025 Official LLM Policy:
neurips.cc/Conferences/...

I really enjoyed "When We Cease to Understand the World", although it's more fiction than history of science

“Science in history” by Bernal is my first recommendation. The work of Ian Hacking is a good recommendation for
Probability

How do task dynamics impact learning in networks with internal dynamics?

Excited to share our ICML Oral paper on learning dynamics in linear RNNs!
with @clementinedomine.bsky.social @mpshanahan.bsky.social and Pedro Mediano

openreview.net/forum?id=KGO...

When we are doing science, we are unknowingly executing our mythology, taken from movies and friends and textbooks, of what science is. History of science helps us ground that myth in reality

I finally wrote a full-fledged blog about this: reading the history of science is an **amazing** yet under-recognized way to develop (emotional) maturity as a researcher.

If you have thoughts/recommendations, please share!
vaishnavh.github.io/2025/04/29/h...

haha that's a new idiom for me. it's perfect! and the flip side is, "target and (potentially) regret", which causes quite a lot of stress. (what if your work gets rejected by the community or worse, overlooked)

but these pressures are real and have always persisted.

</end rant>

I think @abeirami.bsky.social may be interested in this rant.

but now I've the maturity to seek validation from things like "a specific person complimenting my work" or even better, "a meaningful citation where someone substantially builds on my work." (ofc, i also seek internal validation/satisfaction but I gotta be realistic, lol).

i had intense first-hand struggle with a lot of these effects in my phd since i had <= 1 paper/year for the most part. i started managing it only after getting visibly recognized by experts for one of my papers at one point. i still struggle with it at some level.

then there are many other insidious feedback cycles like the fact that publishing more => more visibility => more opportunities/networks/interfaces with the community/more citations => more opportunities/internships etc., => more papers

for example, with the advent of twitter, there's a pressure to stay constantly visible and to have many different things to say every now and then (bec everyone else is doing that), rather than pitch your one paper again and again which starts feeling awkward :-(

someday I hope to write a blog about "all the other forces that discourage me" from publishing less. people always say "publish less!" but without acknowledging these varied and nuanced forces

all other incentivation strategies I had thought of are much more negative/mean. like
- "evaluating someone based on some bottom k papers" or
- "judging negatively for publishing >N papers"

haha thank you! honored you feel that way!

btw, i just noticed this, this sort of a compliment is actually a great way to incentivize people to be more selective in publishing papers (and to counter all the other forces that discourage me from my rate of ~1 paper a year)

Title of paper "Causal Estimation of Tokenisation Bias" and schematic of how we define tokenisation bias, which is the causal effect we are interested in.

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

How does in-context learning emerge in attention models during gradient descent training?

Sharing our new Spotlight paper @icmlconf.bsky.social: Training Dynamics of In-Context Learning in Linear Attention
arxiv.org/abs/2501.16265

Led by Yedi Zhang with @aaditya6284.bsky.social and Peter Latham

This paper is quite nice. It mixes some useful toy models of creativity with insights about how to induce more creativity in LLMs that are better than greedy sampling

Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction We design a suite of minimal algorithmic tasks that are a loose abstraction of open-ended real-world tasks. This allows us to cleanly and controllably quantify the creative limits of the present-day l...

Read the full paper: arxiv.org/abs/2504.15266

This work was with great collaborators at CMU: @chenhenrywu.bsky.social who co-led, Charles Ding & @adtraghunathan.bsky.social! Go follow them to see what else they’re up to! 11/

But, there's a lot of scope for exciting work:
→ generalizing these insights to real cows,
→ studying RL/CoT for creativity,
→ understanding surprising behaviors of seed-conditioning 10/👇🏽

Of course, this is all a study of spherical cows. 🐮
Given the noisy, subjective studies of real cows, we believe an objective study brings
→much-needed clarity of thought (like disentangling the two modes of creativity),
→more ideas,
→better-defined experiments. 9/👇🏽

Our vision is that seed-conditioning can help models sample a latent thought and articulate that one thought into words,

but temp sampling has to articulate multiple latent thoughts in parallel to produce a marginal next-word distribution -- this is more burdensome! 8/👇🏽

Figure showing algorithmic creativity with and without seed-conditioning.

Next, we revisit how to produce randomness: the go-to temp sampling 🌡️ vs. injecting a random prefix (seed-conditioning). 🌱

Remarkably, seed-conditioning produces meaningful diversity even w *greedy* decoding 🤑; it is competitive with temp & in some conditions, superior. 7/👇🏽

…a (deterministic) many-hop planning task.

our work shows inefficiency in tasks
- as simple as 2-hop (stochastic) planning 👶🏽
- where no token-reordering is next-token friendly! 🔀🔃🙅🏾 The model *must* learn all tokens in tandem to detect an *implicit* pattern. 6/👇🏽

Figure from prior paper showing path-star task, a pitfall of next-token prediction.

We argue: creativity requires appreciating big-picture 🖼️ patterns & orchestrating interdependent random decisions in advance ("a leap of thought"). Next-token learning should be inefficient at this.

This complements the known pitfall of next-token learning in the path-star task … 5/👇🏽

Plots report empirical results of algorithmic creativity under next-token vs multi-token objectives.

On these tasks, we can objectively evaluate how “creative” (correct, diverse & original) a model is. 🧑🏽‍🔬

First: Next-token-trained models are largely less creative & memorize much more than multi-token ones (we tried diffusion and teacherless training). 4/👇🏽

Fig 1 screenshot from paper describing minimal algorithmic tasks for combinational creativity (drawing upon different pieces of memory.)

Minimal algorithmic tasks describing exploratory creativity (devising fresh patterns subject to some rules.) Screenshot of Fig 2 in the paper.

Screenshot from poster describing the two tasks in a less technical way.

Our idea was to design minimal, open-ended, graph algorithmic tasks 🧮 abstracting 2 key modes of creativity:

1. Combinational: making surprising connections from memory, like in wordplay.🧠

2. Exploratory: devising fresh patterns obeying some rules, like in problem-design🧩 3/👇🏽