Gary Marcus

Rank 19 of 47
|
Score 64
6/25/2025, 2:57:37 AM
X
In reply to:
Laura Ruis
@LauraRuis
·
18d
Many more interesting findings in the preprint: https://arxiv.org/abs/2506.18777

Awesome work by first authors @JonnyCoook and @silviasapora , and collaborators @aahmadian_ , @akbirkhan , @_rockt , and @j_foerst
Laura Ruis
@LauraRuis
·
18d
We just discovered this and models show limited performance, but we have many follow-up questions. We find big improvements with data and model scaling, so how far can we take this? How many programs can we store in the LLM? Can models internalise things like recursion or search?
Laura Ruis
@LauraRuis
·
18d
We hypothesise that PBB may be an additional reason why code training helps downstream reasoning; models internalise reusable algorithmic abstractions.
Laura Ruis
@LauraRuis
·
18d
Strikingly, using RL in the second stage allows models to evaluate programs seen in an earlier SFT stage (without I/O pairs). By contrast, SFT is much worse at this backward generalisation. Our ablations indicate it's the on-policy nature of RL enabling this.
Laura Ruis
@LauraRuis
·
18d
We also find LLMs can sometimes compose programs only trained on in isolation. Training on foo() and bar() as standalone code snippets allows models to eval baz(x) = foo(bar(x)). GPT-4o can do this w/o CoT, evaluating compositions in-weights (a form of out-of-context reasoning).
Laura Ruis
@LauraRuis
·
18d
We find models can learn an input-general understanding of algorithms from a *single* piece of code. This indicates training LLMs with next-token prediction on source code can overcome (part of) the embers of autoregression (@RTomMcCoy).

The statement engages in a technical discussion about the capabilities and limitations of language models in handling complex tasks, referencing specific research papers. It seeks to relate findings from different studies, promoting a deeper understanding of AI model performance.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does no harm and is focused on academic inquiry. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    It promotes understanding by connecting different research findings. [+2]
  3. Principle 4:
    I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
    It engages constructively by asking questions and seeking clarification. [+1]