Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Kiruluta, Andrew, Lemos, Andreas, Burity, Priscilla
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2506.11108
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909647337684992
author Kiruluta, Andrew
Lemos, Andreas
Burity, Priscilla
author_facet Kiruluta, Andrew
Lemos, Andreas
Burity, Priscilla
contents We present CAGSR-vLLM-MTC, an extension of our Self-Supervised Cross-Attention-Guided Reinforcement (CAGSR) framework, now implemented on the high-performance vLLM runtime, to address both multi-turn dialogue and chain-of-thought reasoning. Building upon our original single-turn approach, we first instrumented vLLM's C++/CUDA kernels to asynchronously capture per-layer, per-head cross-attention weights during generation. We then generalized our self-supervised reward function to accumulate attention signals over entire conversation histories and intermediate chain-of-thought steps. We discuss practical trade-offs, including an entropy-based clamping mechanism to prevent attention collapse on early context, and outline future directions for multi-party dialogues and hierarchical reasoning.
format Preprint
id arxiv_https___arxiv_org_abs_2506_11108
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle History-Aware Cross-Attention Reinforcement: Self-Supervised Multi Turn and Chain-of-Thought Fine-Tuning with vLLM
Kiruluta, Andrew
Lemos, Andreas
Burity, Priscilla
Computation and Language
We present CAGSR-vLLM-MTC, an extension of our Self-Supervised Cross-Attention-Guided Reinforcement (CAGSR) framework, now implemented on the high-performance vLLM runtime, to address both multi-turn dialogue and chain-of-thought reasoning. Building upon our original single-turn approach, we first instrumented vLLM's C++/CUDA kernels to asynchronously capture per-layer, per-head cross-attention weights during generation. We then generalized our self-supervised reward function to accumulate attention signals over entire conversation histories and intermediate chain-of-thought steps. We discuss practical trade-offs, including an entropy-based clamping mechanism to prevent attention collapse on early context, and outline future directions for multi-party dialogues and hierarchical reasoning.
title History-Aware Cross-Attention Reinforcement: Self-Supervised Multi Turn and Chain-of-Thought Fine-Tuning with vLLM
topic Computation and Language
url https://arxiv.org/abs/2506.11108