arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2603.01375cs.AIcs.LG
收藏

词语与权重:通过协同适应流式多轮交互

Words & Weights: Streamlining Multi-Turn Interactions via Co-Adaptation

Shenzhen University(深圳大学);Hong Kong University of Science(香港科学大学);Guangdong Laboratory of Artificial Intelligence(广东人工智能与数字经济实验室);University of Science(科学大学);The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳));Carleton University(卡尔顿大学)

Chenxing Wei, Hong Wang, Ying He, Zhongxiang Dai, Bo Jiang, F. Richard Yu, Yao Shu

AI总结:

ROSA2通过协同适应词语与权重空间,提升多轮交互的语义清晰度和参数更新效率,实现更高效的交互性能。

摘要:

在推理时间对多轮交互进行策略适应(T2PAM)对于使大型语言模型(LLMs)与动态用户需求对齐至关重要。然而,现有方法通常将测试时间适应视为单一轴问题,要么纯粹优化指令(提示工程),要么仅调整权重(测试时间训练),忽视了交互失败源于模糊性和能力不足的耦合混合。我们主张这两种优化路径并非简单相加,而是协同作用:语义清晰性是有效参数更新的先决条件。为此,我们提出了ROSA2框架,将交互重新表述为一个联合优化问题,涉及词语和权重的异质空间。通过数学分解误差信号,ROSA2利用文本梯度纠正意图模糊性,并利用参数更新弥补能力差距。理论上,我们证明这种协同适应严格减少了收敛所需的参数偏移。实验表明,ROSA2在MATH上比最先进的基线高出30%,同时将交互轮次减少40%,证明优化上下文解锁了参数更新的真正潜力。

英文摘要:

Test-time policy adaptation for multi-turn interactions (T2PAM) is essential for aligning Large Language Models (LLMs) with dynamic user needs during inference time. However, existing paradigms commonly treat test-time adaptation as a single-axis problem, either purely refining instructions (Prompt Engineering) or only adjusting weights (Test-Time Training), ignoring that interaction failures stem from a coupled mix of ambiguity and incapacity. We argue that these two optimization paths are not merely additive but synergistic: semantic clarity acts as a pre-conditioner for effective parameter updates. To this end, we propose ROSA2, a framework that reformulates interaction as a joint optimization problem over the heterogeneous space of Words and Weights. By mathematically decomposing the error signal, ROSA2 utilizes textual gradients to rectify intent ambiguity and parameter updates to bridge capability gaps. Theoretically, we prove that this co-adaptation strictly reduces the required parameter shift for convergence. Empirically, ROSA2 outperforms state-of-the-art baselines by 30% on MATH while reducing interaction turns by 40%, demonstrating that refining the context unlocks the true potential of parameter updates.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑