Output-token budgets are a first-order deployment constraint for large language model (LLM) agents, yet existing benchmarks typically evaluate them under a single fixed or generous budget. We introduce TokenCliff, a benchmark for measuring multi-turn agent performance across episode-level output-token budgets. It contains 660 deterministic tasks spanning four capability-oriented domains—file operations, data transformation, tool use, and planning—plus a compositional multi-step split. We propose two budget-sensitivity metrics: Budget Elasticity, which captures robustness across budget levels, and the Cliff Index, which captures the sharpest adjacent-budget performance drop. Evaluating 10 LLMs at five budget levels from 200 to 8,192 tokens yields 33,000 episodes. Results show that (1) accuracy drops by up to 32 percentage points between adjacent budgets; (2) neither peak performance nor larger model size necessarily implies stronger budget elasticity; and (3) budget sensitivity is substantially greater in the evaluated multi-turn setting than in the single-shot diagnostic. Code and data are available in our public repository: https://github.com/BairuiZhang/TokenCliff

Legal article retrieval is critical for building traceable and reliable legal AI systems, where conclusions must be grounded in specific legal articles. However, existing open-domain retrieval methods rely heavily on surface-level lexical or semantic similarity, making it difficult for them to distinguish legally relevant articles from those that are textually similar but legally inapplicable or misaligned with the user’s underlying intent. To bridge this gap, we propose **LexPath**, a domain-oriented multi-path framework comprising a multi-path retrieval module and an intent-aware reranking module. The retrieval module combines two complementary legal-specific paths to collect candidate articles: an IRAC-guided sparse path that expands queries with legally informative keywords, and a structure-guided dense path trained with hard negatives derived from legal hierarchy and citation relations. The reranking module then further refines the candidate ranking by incorporating the intent consistency score between queries and legal articles. We evaluate **LexPath** on three Chinese benchmarks spanning diverse query scenarios, and further assess its cross-jurisdictional applicability on a Japanese benchmark. Experimental results demonstrate that LexPath consistently outperforms lexical, dense, hybrid, and adaptive retrieval-augmented generation (RAG) baselines across the four benchmarks. Ablation studies further verify the effectiveness of each component. The data and code are available at: <https://github.com/wxliuabigail/LexPath>.