Task-decomposition reranking beats the best baseline 78.7 vs 73.1 on ALFWorld-unseen, using just 1.3 skills per task

Notes on Task Decomposition-Guided Reranking for Adaptive Agent Skill Retrieval (arXiv:2607.06283) — Yanping Chen, Weijie Shi, Wen Yang, Jiajie Xu · July 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SkillReranker treats skill selection as a graph-matching problem, not a similarity search. It decomposes a task into an ordered sequence of subtasks and intermediate sub-states, then parses every candidate skill the same way: a precondition state — what has to be true before it fires — and a completion state — what's true once it runs. Those states become nodes in a directed acyclic execution graph, with each skill wired in as an edge from whichever node matches its precondition to whichever downstream node matches its completion. That structure is a direct answer to what the paper calls the granularity gap: task requirements are specific, skill descriptions are written generic for reuse, and flat semantic matching keeps mistaking topical overlap for functional fit. Once the graph exists, SkillReranker walks it node by node, applying a split condition — comparing the strength of skills that complete before a node, start after it, or merely pass across it — to carve the task into intervals. Inside each interval, a cross-encoder scores the surviving candidates by summing the log-probability of two relevance checks, one against the full task instruction and one against just that interval's subtask text (equivalent to their geometric mean), and keeps the top scorer.

Across ALFWorld and ScienceWorld, tested with three backbone LLMs — DeepSeek-v4-Flash, GPT-5.4-Mini, Qwen3.6-27B — against a generative LLM-as-selector, the fine-tuned SkillRouter, and the dependency-graph retriever Graph of Skills, SkillReranker won 11 of 12 reward comparisons and 11 of 12 step-count comparisons. On ALFWorld's unseen split with DeepSeek-v4-Flash, reward climbed from 73.14 (SkillRouter, the strongest baseline) to 78.73, with fewer environment-interaction steps and lower token spend along the way. The 'adaptive' half of the design shows up in a single number: SkillReranker's own average skill-set size sits at 1.28–1.30 skills per task across all four seen/unseen splits. A fixed top-k retriever can't produce that — it's the graph's interval count deciding how many skills a given task actually needs. An ablation isolates the load-bearing piece: removing the structured-parsing step costs more reward than removing the graph edges or the split logic, and the weaker Qwen3.6-27B backbone takes the biggest hit, its ALFWorld-seen reward falling from 78.57 to 65.00 without it.