paper-with-me

Papers

An empirical study on the limitation of Transformers in program trace generation

2025-09-29 · Simeng Sun arxiv

We study Transformers on the task \emph{program trace generation} (PTG), where models produce step-by-step execution traces for synthetic programs. Unlike existing algorithmic problems, PTG externalizes reasoning through long traces where each step is trivial. We train small Transformers with diverse modifications, including alternative position encodings, softmax replacements, hybrid model, and short convolutions. While these models achieve strong in-distribution accuracy, they exhibit systematic failures when generalizing to various factors (e.g., program length, trace steps), though some designs significantly improve generalization.

📄 PDF Abstract BibTeX arXiv:2509.25073

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

Can Transformers Reason Logically? A Study in SAT Solving

2024-10-09 · Leyan Pan, Vijay Ganesh, Jacob Abernethy, Chris Esposo 외

We formally study the logical reasoning capabilities of decoder-only Transformers in the context of the boolean satisfiability (SAT) problem. First, we prove by construction that decoder-only Transformers can decide 3-SA…

DecoderLogical ReasoningOut-of-Distribution Generalization

Debugging code world models

2026-02-07 · Babak Rahmani arxiv

Code World Models (CWMs) are language models trained to simulate program execution by predicting explicit runtime state after every executed command. This execution-based world modeling enables internal verification with…

Knee-Deep in C-RASP: A Transformer Depth Hierarchy

2025-06-19 · Andy Yang, Michaël Cadilhac, David Chiang

It has been observed that transformers with greater depth (that is, more layers) have more capabilities, but can we establish formally which capabilities are gained with greater depth? We answer this question with a theo…

On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: A Shortest-Path Case Study

2025-07-07 · Riccardo Alberghi, Elizaveta Demyanenko, Luca Biggio, Luca Saglietti arxiv

Recent advances in natural language processing highlight two key factors for improving reasoning in large language models (LLMs): (i) allocating more test-time compute tends to help on harder problems but often introduce…

Exploiting ftrace's function_graph Tracer Features for Machine Learning: A Case Study on Encryption Detection

2025-12-04 · Kenan Begovic, Abdulaziz Al-Ali, Qutaibah Malluhi arxiv

This paper proposes using the Linux kernel ftrace framework, particularly the function graph tracer, to generate informative system level data for machine learning (ML) applications. Experiments on a real world encryptio…

Anomaly Detection