Automata from Agent Traces: Failure and Next-Step Prediction
为解决LLM代理行为结构不透明问题，本文提出将多条轨迹压缩成一个紧凑的有限状态机(FSM)，用于预测下一步行动和失败。
0 citationsRead paper
Selected work
Representative Papers
AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace
AgentRoom通过CRDT支持的共享工作空间解决并发多代理编码中的协调问题，减少任务放弃率并提高稳定性。
0 citationsRead paper
Routing Is Least Learnable Where It Is Most Valuable: Bounds on Representation Routing for Web Agents
This study investigates dynamic selection of optimal observation modalities—text, pixels, or their fusion—for web automation tasks to balance success rate and computational cost. Through systematic evaluation across six observation modalities on diverse website–model combinations, the work proposes task-adaptive routing strategies, including confidence cascades, cost-aware tiering, and zero-cost rules, and uncovers a negative feedback loop between the scarcity of routing supervision signals and task difficulty. Experimental results demonstrate that a fixed, high-quality observation modality often outperforms complex routing schemes; furthermore, routing only failed tasks to the cheapest modality reduces costs by 9.5–30.6% without compromising success rates. The authors publicly release a complete measurement protocol and re-execution noise intervals, establishing clear boundaries for effective routing under cost constraints.
0 citationsRead paper
Quantization Damage Is Multiplicative, Not Additive
This work addresses the decision-silencing failure induced by low-bit quantization in large language models, a phenomenon poorly predictable by existing methods. By analyzing shifts in confidence margins on binary-choice tasks before and after quantization, the study reveals— for the first time—that quantization damage manifests as multiplicative compression of decision boundaries rather than additive noise. The authors propose a “boundary contraction” mechanism to explain the directionality of decision failures. Through extensive experiments spanning 16 models, three quantization schemes, and bit-widths from 2 to 8 bits, combined with statistical fitting and calibration evaluation, their approach accurately predicts retained decision flip rates (median error of only 1.8 percentage points) and achieves a remarkably low calibration error of 0.004. Furthermore, they demonstrate that adding just one bit is more cost-effective than current repair strategies.
0 citationsRead paper
When Should Graph Attention Be Sparse? Learning a Per-Edge Tsallis Index
This work addresses the limitation of conventional graph attention mechanisms, which rely on fixed softmax normalization and struggle to simultaneously accommodate the divergent attention requirements of homophilic and heterophilic graph structures. To overcome this, the authors propose the Learnable Tsallis Graph Attention (LTGA) layer, which introduces a learnable Tsallis entropy index \( q \) into the attention mechanism for the first time. This enables dynamic, edge-wise modulation of attention distributions, allowing continuous interpolation between heavy-tailed and compactly supported forms. The method facilitates edge-level sparsification and interpretable attention pruning, and is end-to-end trainable via a Tsallis entropy–based differentiable reparameterization. Evaluated across eight benchmark datasets, LTGA achieves the best average rank (2.75), automatically zeroes out 42% of erroneous edge attentions, and ablating these pruned edges results in a 7.1-point performance drop—significantly outperforming random pruning.
0 citationsRead paper
Recent publications
Latest Papers
Automata from Agent Traces: Failure and Next-Step Prediction
为解决LLM代理行为结构不透明问题，本文提出将多条轨迹压缩成一个紧凑的有限状态机(FSM)，用于预测下一步行动和失败。
0 citationsRead paper
AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace
AgentRoom通过CRDT支持的共享工作空间解决并发多代理编码中的协调问题，减少任务放弃率并提高稳定性。
0 citationsRead paper
Routing Is Least Learnable Where It Is Most Valuable: Bounds on Representation Routing for Web Agents
This study investigates dynamic selection of optimal observation modalities—text, pixels, or their fusion—for web automation tasks to balance success rate and computational cost. Through systematic evaluation across six observation modalities on diverse website–model combinations, the work proposes task-adaptive routing strategies, including confidence cascades, cost-aware tiering, and zero-cost rules, and uncovers a negative feedback loop between the scarcity of routing supervision signals and task difficulty. Experimental results demonstrate that a fixed, high-quality observation modality often outperforms complex routing schemes; furthermore, routing only failed tasks to the cheapest modality reduces costs by 9.5–30.6% without compromising success rates. The authors publicly release a complete measurement protocol and re-execution noise intervals, establishing clear boundaries for effective routing under cost constraints.
0 citationsRead paper
Quantization Damage Is Multiplicative, Not Additive
This work addresses the decision-silencing failure induced by low-bit quantization in large language models, a phenomenon poorly predictable by existing methods. By analyzing shifts in confidence margins on binary-choice tasks before and after quantization, the study reveals— for the first time—that quantization damage manifests as multiplicative compression of decision boundaries rather than additive noise. The authors propose a “boundary contraction” mechanism to explain the directionality of decision failures. Through extensive experiments spanning 16 models, three quantization schemes, and bit-widths from 2 to 8 bits, combined with statistical fitting and calibration evaluation, their approach accurately predicts retained decision flip rates (median error of only 1.8 percentage points) and achieves a remarkably low calibration error of 0.004. Furthermore, they demonstrate that adding just one bit is more cost-effective than current repair strategies.
0 citationsRead paper
When Should Graph Attention Be Sparse? Learning a Per-Edge Tsallis Index
This work addresses the limitation of conventional graph attention mechanisms, which rely on fixed softmax normalization and struggle to simultaneously accommodate the divergent attention requirements of homophilic and heterophilic graph structures. To overcome this, the authors propose the Learnable Tsallis Graph Attention (LTGA) layer, which introduces a learnable Tsallis entropy index \( q \) into the attention mechanism for the first time. This enables dynamic, edge-wise modulation of attention distributions, allowing continuous interpolation between heavy-tailed and compactly supported forms. The method facilitates edge-level sparsification and interpretable attention pruning, and is end-to-end trainable via a Tsallis entropy–based differentiable reparameterization. Evaluated across eight benchmark datasets, LTGA achieves the best average rank (2.75), automatically zeroes out 42% of erroneous edge attentions, and ablating these pruned edges results in a 7.1-point performance drop—significantly outperforming random pruning.
0 citationsRead paper