SkillSight: Seeing Through Shared Descriptions for Accurate Skill Retrieval

# SkillSight: Seeing Through Shared Descriptions for Accurate Skill Retrieval
Source: [https://arxiv.org/html/2607.18785](https://arxiv.org/html/2607.18785)
Jinying Xiao1, Bin Ji1, Shasha Li1, Xiaodong Liu1, Ma Jun1, Jiacheng Jie1, Chao Wang2, Nyima Tashi3, Jie Yu1

###### Abstract

As large language model agents gain access to increasingly large skill libraries, retrieving the right skill becomes critical to reliable capability selection and execution\. Existing retrievers often treat skill descriptions as ordinary documents, overlooking their highly regular structure: shared descriptive patterns recur across many skills while providing little evidence for distinguishing the required capability\. We show that this shared descriptive background systematically contributes to dense relevance scores, induces a pronounced energy gap between queries and skill documents, and obscures task\-relevant signals\. Based on this observation, we propose SkillSight, a training\-free retrieval framework that calibrates shared background in both semantic and lexical spaces\. Semantic Background Calibration estimates a background subspace from generic tokens identified by IDF, reducing similarity induced by shared descriptive patterns, while Lexical Evidence Calibration downweights shared background tokens to recover discriminative token\-level evidence\. Experiments on SRA\-Bench and SkillBench\-Supp demonstrate consistent improvements across retrieval metrics, with SkillSight improving Recall@10 by up to 20\.21 percentage points over the original dense retriever\. In end\-to\-end evaluation, SkillSight achieves the best overall performance across three agent models and outperforms LLM Selection by up to 4\.97 percentage points\. It is also up to 1,248 times faster than the Dense \+ Reranker baseline\. These results identify shared descriptive background as a key source of bias in skill retrieval and demonstrate that explicitly calibrating it enables accurate and efficient skill selection without additional training\. Our code is available at https://github\.com/xiaojinying/SkillSight\.

## Introduction

As large language model agents gain access to an increasing number of skills\(Xuet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib3); Dinget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib2)\), accurately identifying the capabilities required for a given task from a large skill repository becomes a critical retrieval problem in agent systems\. A skill document typically specifies the capability scope, invocation interface, execution conditions, and usage instructions\(Yuanet al\.[2025](https://arxiv.org/html/2607.18785#bib.bib11)\)\. Providing all documents directly to a large language model incurs substantial context overhead and increases the risk of selecting or invoking irrelevant capabilities\(Qinet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib1)\)\. Skill retrieval addresses this problem by returning a compact set of relevant candidates\. The agent then selects and executes the appropriate skill\. Retrieval quality therefore directly affects the agent’s ability to use external capabilities effectively\.

Existing studies improve skill retrieval through query reformulation, specialized retrievers, candidate reranking, and tool\-relation modeling\(Chenet al\.[2024b](https://arxiv.org/html/2607.18785#bib.bib36); Zhenget al\.[2024](https://arxiv.org/html/2607.18785#bib.bib37); Qinet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib1); Quet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib34)\), primarily addressing query–skill semantic mismatch or reassessing candidate relevance with stronger models\. However, skill retrieval exhibits a distinctive input asymmetry: queries express instance\-specific objectives, targets, and constraints, whereas reusable skill documents follow regular descriptive structures and extensively share content about capability specifications, invocation conditions, and execution procedures\(Lianget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib32)\)\. This structural regularity can induce document\-side background bias in dense retrieval, because query–document similarity aggregates task\-relevant evidence with corpus\-shared descriptive content\. Consequently, a hard negative may receive a high score by matching generic interfaces or usage instructions despite mismatching the required operation, target, or constraints\. Existing representation correction methods remove common embedding directions\(Mu and Viswanath[2018](https://arxiv.org/html/2607.18785#bib.bib8); Zhouet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib6)\), while lexical augmentation and interaction\-based reranking introduce complementary matching signals\(Bruchet al\.[2023](https://arxiv.org/html/2607.18785#bib.bib18); Zhuanget al\.[2024](https://arxiv.org/html/2607.18785#bib.bib19)\); however, neither explicitly identifies shared skill descriptions as a semantic source of ranking bias nor quantifies their contribution to relevance scores\.

We analyze this problem from three perspectives: ranking behavior, token statistics, and representation geometry\. First, compared with conventional text retrieval datasets\(Thakuret al\.[2021](https://arxiv.org/html/2607.18785#bib.bib14)\), skill retrieval datasets require a greater retrieval depth to achieve the same coverage of gold skills, and a larger proportion of candidate documents are ranked ahead of the gold skill\. Second, the gold\-skill rates of generic tokens that frequently occur across skill documents are concentrated near zero, indicating that these tokens provide limited discrimination among skills\. We further use generic tokens to estimate the background subspace of a skill corpus and observe that skill documents consistently exhibit higher background energy than queries in this subspace\. This query–document energy gap is not evident in conventional text retrieval datasets\. An orthogonal decomposition of the dense score further shows that background alignment contributes additional scores to some candidates, thereby reducing the influence of task\-specific evidence on ranking\.

Based on these findings, we propose SkillSight, a training\-free relevance calibration framework for skill retrieval\. Semantic Background Calibration \(SBC\) estimates a corpus\-level background subspace from generic tokens and removes its contribution to dense similarity, while Lexical Evidence Calibration \(LEC\) downweights corpus\-common terms to emphasize fine\-grained capability\-matching evidence\. The calibrated semantic and lexical scores are fused within a local candidate set without additional training or model inference\.

Experiments on SRA\-Bench and SkillBench\-Supp show that SkillSight improves Recall@10 by up to 20\.21 percentage points over the original dense retriever\. It also achieves the best non\-oracle end\-to\-end performance across three agent models, outperforming LLM Selection by up to 4\.97 points, while being up to 1,248×\\timesfaster than neural reranking\.

Our contributions are as follows:

- •We identify and systematically analyze document\-side background bias in skill retrieval\. Through generic\-token statistics, background energy analysis, and dense\-score decomposition, we reveal how shared descriptive structures affect relevance ranking\.
- •We propose SkillSight, which suppresses the contribution of shared background to dense relevance scores through Semantic Background Calibration and recovers fine\-grained capability\-matching evidence through Lexical Evidence Calibration\. The proposed framework requires neither additional training nor model inference\.
- •We conduct comprehensive evaluations on two skill retrieval benchmarks and three agent models\. The results demonstrate the advantages of SkillSight in offline retrieval, end\-to\-end task performance, and inference efficiency, while ablation studies confirm that SBC and LEC provide effective and complementary relevance signals\.

## Related Work

### Agent Tool and Skill Retrieval

As tool libraries grow, exposing all descriptions to an LLM incurs substantial context overhead and increases selection errors\. ToolLLM introduces a dedicated API retriever\(Qinet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib1)\), while subsequent work improves retrieval through query reformulation, reranking, and tool\-relation modeling\(Chenet al\.[2024b](https://arxiv.org/html/2607.18785#bib.bib36); Zhenget al\.[2024](https://arxiv.org/html/2607.18785#bib.bib37); Quet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib34)\)\. Recent studies extend this line to skill retrieval and agent execution\(Shiet al\.[2025](https://arxiv.org/html/2607.18785#bib.bib29); Liet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib33); Suet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib28)\)\. SkillRouter\(Zhenget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib23)\)and SkillRet\(Choet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib7)\)both train specialized two\-stage embedding and reranking pipelines\.

GoS\(Liuet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib31)\)and GoSkills\(Zenget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib30)\)first use LLM\- and encoder\-assisted relation induction to transform a flat skill library into a structured graph, from which they retrieve dependency\-aware bundles or role\-labeled groups\. This changes the retrieval corpus and introduces additional graph\-construction costs, making them not directly comparable to our setting, which ranks individual skills from the original library without graph preprocessing\.

### Retrieval Representations and Relevance Calibration

Dense retrieval relies on the overall semantic similarity between query and document representations\(Karpukhinet al\.[2020](https://arxiv.org/html/2607.18785#bib.bib12)\), but may underemphasize fine\-grained evidence such as operations, entities, and constraints\. Existing hybrid retrieval methods address this limitation by combining dense and lexical signals\(Zhuanget al\.[2024](https://arxiv.org/html/2607.18785#bib.bib19); Bruchet al\.[2023](https://arxiv.org/html/2607.18785#bib.bib18)\)\. Another line of work reduces the influence of non\-discriminative components on similarity estimation by removing common directions from representations or improving the geometry of the embedding space\(Aroraet al\.[2017](https://arxiv.org/html/2607.18785#bib.bib17); Mu and Viswanath[2018](https://arxiv.org/html/2607.18785#bib.bib8); Zhouet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib6)\)\. These methods typically identify common components from the global representation distribution without explicitly characterizing their task\-specific semantic origins\.

## Analysis

### Shared Descriptive Background

A dense retriever uses the inner product between the query and document embeddings as the relevance score:

s​\(q,d\)=𝐪⊤​𝐝\.s\(q,d\)=\\mathbf\{q\}^\{\\top\}\\mathbf\{d\}\.\(1\)This score aggregates all alignment signals in the embedding space\. In skill retrieval, these signals can be broadly divided into two categories\. The first captures task\-specific evidence required by the query, whereas the second reflects the explanatory background shared across skill documents\. The former determines whether a skill actually satisfies the current request, while the latter mainly indicates whether a document follows a typical skill\-description pattern\.

![Refer to caption](https://arxiv.org/html/2607.18785v1/x1.png)Figure 1:Retrieval ranking behavior for SRA\-Bench and ordinary\-text benchmarks using Qwen3\-Embedding\-0\.6B with cosine similarity\. \(a\) Gold recall over Top\-k retrieval depth on a log scale, where gold recall is the fraction of queries whose best gold document appears within the top k results\. \(b\) Average percentage of corpus documents ranked above the best gold document, computed as mean\(\(r​a​n​k−1\)/c​o​r​p​u​s​s​i​z​e\)×100\(\(rank\-1\)/corpus\\ size\)\\times 100\.This mixture of signals affects the ranking of hard negatives\. Figure[1](https://arxiv.org/html/2607.18785#Sx3.F1)shows that the recall curve for gold skills on SRA\-Bench consistently lags behind that observed in conventional text retrieval\(Thakuret al\.[2021](https://arxiv.org/html/2607.18785#bib.bib14)\), with a larger average proportion of candidate documents ranked ahead of the gold document\. This result indicates that the dense score reflects not only the capabilities required by the task, but also systematic document\-side background similarity\. Specifically, a hard negative may resemble the gold skill in writing style, interface description, and semantic structure, thereby receiving a high dense score, even though it mismatches the query in key operations, target entities, or domain conditions\. Such background alignment weakens the ranking contribution of fine\-grained capability\-matching signals and introduces a systematic bias into skill retrieval\.

This observation motivates us to first identify the shared background structure in the skill corpus and then examine how it contributes to the dense score\. An ideal relevance score should preserve the semantic content of skill documents while limiting the excessive contribution of background signals, so that the resulting ranking depends more strongly on query\-specific evidence\.

![Refer to caption](https://arxiv.org/html/2607.18785v1/x2.png)Figure 2:Gold\-skill rate distributions of different token groups on SRA\-Bench and SkillBench\-Supp\. The horizontal axis represents the proportion of skills containing a given token that serve as the gold skill, while the vertical axis shows the corresponding number of tokens on a log scale\.
### Query–Document Background Energy Gap

We characterize the shared background of a skill corpus through token commonality\. Let𝒟=\{di\}i=1N\\mathcal\{D\}=\\\{d\_\{i\}\\\}\_\{i=1\}^\{N\}denote the skill documents with normalized embeddings\{𝐝i\}i=1N\\\{\\mathbf\{d\}\_\{i\}\\\}\_\{i=1\}^\{N\}\. For each corpus tokentt, we define

idf⁡\(t\)=log⁡N\+1df⁡\(t\)\+1\+1,\\operatorname\{idf\}\(t\)=\\log\\frac\{N\+1\}\{\\operatorname\{df\}\(t\)\+1\}\+1,\(2\)wheredf⁡\(t\)\\operatorname\{df\}\(t\)is the number of documents containingtt\. We apply Otsu’s method\(Otsu[1979](https://arxiv.org/html/2607.18785#bib.bib13)\)to the IDF distribution and denote the resulting low\-IDF token set by𝒯g\\mathcal\{T\}\_\{g\}\. Figure[2](https://arxiv.org/html/2607.18785#Sx3.F2)shows that the gold\-skill rates of these generic tokens concentrate near zero, whereas other tokens retain greater mass in higher\-rate regions\. This indicates that corpus\-frequent tokens primarily encode shared descriptions and provide limited discrimination among skills\.

We map this token\-level commonality into the embedding space\. For eacht∈𝒯gt\\in\\mathcal\{T\}\_\{g\}, letℐt=\{i∣t∈di\}\\mathcal\{I\}\_\{t\}=\\\{i\\mid t\\in d\_\{i\}\\\}and compute its associated mean document embedding:

𝐠t=1\|ℐt\|​∑i∈ℐt𝐝i\.\\mathbf\{g\}\_\{t\}=\\frac\{1\}\{\|\\mathcal\{I\}\_\{t\}\|\}\\sum\_\{i\\in\\mathcal\{I\}\_\{t\}\}\\mathbf\{d\}\_\{i\}\.\(3\)We collect these token\-induced directions together with the corpus mean𝐝¯=1N​∑i=1N𝐝i\\bar\{\\mathbf\{d\}\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbf\{d\}\_\{i\}and perform singular value decomposition:

G=\[𝐠t1,…,𝐠tm,𝐝¯\]=U​Σ​V⊤,B=\[𝐮1,…,𝐮r\],G=\\left\[\\mathbf\{g\}\_\{t\_\{1\}\},\\ldots,\\mathbf\{g\}\_\{t\_\{m\}\},\\bar\{\\mathbf\{d\}\}\\right\]=U\\Sigma V^\{\\top\},\\quad B=\[\\mathbf\{u\}\_\{1\},\\ldots,\\mathbf\{u\}\_\{r\}\],\(4\)where\{t1,…,tm\}⊆𝒯g\\\{t\_\{1\},\\ldots,t\_\{m\}\\\}\\subseteq\\mathcal\{T\}\_\{g\}, andrrdenotes the selected rank of the background subspace\. The orthonormal basisBBspans a low\-dimensional subspace that summarizes the background directions shared across skill documents\.

For any normalized embedding𝐯\\mathbf\{v\}, we define its background energy as

eB​\(𝐯\)=‖B⊤​𝐯‖22\.e\_\{B\}\(\\mathbf\{v\}\)=\\left\\\|B^\{\\top\}\\mathbf\{v\}\\right\\\|\_\{2\}^\{2\}\.\(5\)Given normalized query embeddings\{𝐪j\}j=1M\\\{\\mathbf\{q\}\_\{j\}\\\}\_\{j=1\}^\{M\}, the average document\- and query\-side energies are

ED=1N​∑i=1NeB​\(𝐝i\),EQ=1M​∑j=1MeB​\(𝐪j\)\.E\_\{D\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}e\_\{B\}\(\\mathbf\{d\}\_\{i\}\),\\qquad E\_\{Q\}=\\frac\{1\}\{M\}\\sum\_\{j=1\}^\{M\}e\_\{B\}\(\\mathbf\{q\}\_\{j\}\)\.\(6\)
As shown in Figure[3](https://arxiv.org/html/2607.18785#Sx3.F3), conventional text retrieval datasets lie close to the symmetry lineED=EQE\_\{D\}=E\_\{Q\}, whereas skill retrieval datasets consistently exhibitED\>EQE\_\{D\}\>E\_\{Q\}\. Their document\-level energy distributions also shift toward higher\-energy regions, showing that stronger coupling to the background subspace is a systematic property of skill documents\.

This energy gap directly affects dense relevance scoring\. LetPB=B​B⊤P\_\{B\}=BB^\{\\top\}be the projection matrix onto the background subspace\. SincePBP\_\{B\}andI−PBI\-P\_\{B\}project onto orthogonal subspaces, the dense score decomposes as

𝐪⊤​𝐝=\(PB​𝐪\)⊤​\(PB​𝐝\)⏟background alignment\+\(\(I−PB\)​𝐪\)⊤​\(\(I−PB\)​𝐝\)⏟residual alignment\.\\mathbf\{q\}^\{\\top\}\\mathbf\{d\}=\\underbrace\{\(P\_\{B\}\\mathbf\{q\}\)^\{\\top\}\(P\_\{B\}\\mathbf\{d\}\)\}\_\{\\text\{background alignment\}\}\+\\underbrace\{\\bigl\(\(I\-P\_\{B\}\)\\mathbf\{q\}\\bigr\)^\{\\top\}\\bigl\(\(I\-P\_\{B\}\)\\mathbf\{d\}\\bigr\)\}\_\{\\text\{residual alignment\}\}\.\(7\)When documents exhibit elevated background energy, the first term can provide additional scores to candidates whose shared descriptions align with the query\. Because this contribution is only weakly related to the required capability, it can obscure task\-specific evidence and bias the resulting ranking\.

![Refer to caption](https://arxiv.org/html/2607.18785v1/x3.png)Figure 3:Background energy statistics for queries and documents\. \(a\) Average background energy across datasets\. The horizontal axis denotes document energy, and the vertical axis denotes query energy\. The dashed line indicates equal query and document energy, while different markers distinguish skill datasets from ordinary text corpora\. \(b\) Distributions of background energy for individual queries and documents\. The horizontal axis denotes background energy, and the vertical axis denotes density\. Different curves represent the queries and documents in SciFact and ToolQA\.

## SkillSight

The preceding analysis shows that descriptive patterns widely shared across skill documents contribute to the dense relevance score and create a pronounced energy gap between queries and documents\. Based on this observation, SkillSight calibrates shared corpus information at both the semantic representation and lexical matching levels\. The semantic channel removes alignment within the background subspace, while the lexical channel reduces the matching contribution of frequently shared tokens\. SkillSight then combines these two complementary sources of relevance evidence within the candidate set\.

### Semantic Background Calibration

Given an orthonormal basisBBof the background subspace, its projection matrix isPB=B​B⊤P\_\{B\}=BB^\{\\top\}\. SkillSight projects both the query and document representations onto the orthogonal complement of the background subspace:

𝐪⟂=\(I−PB\)​𝐪,𝐝⟂=\(I−PB\)​𝐝\.\\mathbf\{q\}\_\{\\perp\}=\(I\-P\_\{B\}\)\\mathbf\{q\},\\qquad\\mathbf\{d\}\_\{\\perp\}=\(I\-P\_\{B\}\)\\mathbf\{d\}\.\(8\)
The background\-calibrated dense relevance score is defined as

s⟂​\(q,d\)=𝐪⟂⊤​𝐝⟂\.s\_\{\\perp\}\(q,d\)=\\mathbf\{q\}\_\{\\perp\}^\{\\top\}\\mathbf\{d\}\_\{\\perp\}\.\(9\)
According to Eq\. \([7](https://arxiv.org/html/2607.18785#Sx3.E7)\),s⟂​\(q,d\)s\_\{\\perp\}\(q,d\)corresponds to the semantic alignment outside the background subspace and therefore excludes the direct contribution of background alignment\. We do not renormalize the representations after projection, allowing their norms to preserve the residual semantic energy of the original embeddings\. For a document with high background energy, the remaining component available for relevance matching becomes correspondingly smaller, which reduces the influence of shared descriptive patterns on the ranking\.

After background calibration, the dense score provides a more task\-specific estimate of semantic relevance\. Figure[2](https://arxiv.org/html/2607.18785#Sx3.F2)further shows that other tokens retain substantial mass in regions with high gold\-skill fractions, suggesting that lexical signals contain additional discriminative information\. SkillSight therefore incorporates lexical evidence to improve the distinction between specific operations, entities, and constraints\.

### Lexical Evidence Calibration

We first remove the generic tokens from the query token set:

𝒯qsp=𝒯q∖𝒯g\.\\mathcal\{T\}\_\{q\}^\{\\mathrm\{sp\}\}=\\mathcal\{T\}\_\{q\}\\setminus\\mathcal\{T\}\_\{g\}\.\(10\)As shown in Figure[2](https://arxiv.org/html/2607.18785#Sx3.F2), the remaining tokens still vary substantially in discriminative power\. We quantify their residual commonality by document prevalence,

p​\(t\)=df​\(t\)N,p\(t\)=\\frac\{\\mathrm\{df\}\(t\)\}\{N\},\(11\)and assign eacht∈𝒯qspt\\in\\mathcal\{T\}\_\{q\}^\{\\mathrm\{sp\}\}the calibrated weight

wβ​\(t\)=idf​\(t\)​\(1−p​\(t\)\)β,β≥0\.w\_\{\\beta\}\(t\)=\\mathrm\{idf\}\(t\)\\bigl\(1\-p\(t\)\\bigr\)^\{\\beta\},\\qquad\\beta\\geq 0\.\(12\)The factor\(1−p​\(t\)\)β\(1\-p\(t\)\)^\{\\beta\}further suppresses tokens shared by many documents, withβ\\betacontrolling the calibration strength\. Whenβ=0\\beta=0, the weight reduces to standard IDF\.

For a candidate documentdd, we define the lexical relevance score as the weighted coverage of non\-generic query tokens:

slex​\(q,d\)=∑t∈𝒯qspwβ​\(t\)​𝕀​\[t∈d\]∑t∈𝒯qspwβ​\(t\)\.s\_\{\\mathrm\{lex\}\}\(q,d\)=\\frac\{\\displaystyle\\sum\_\{t\\in\\mathcal\{T\}\_\{q\}^\{\\mathrm\{sp\}\}\}w\_\{\\beta\}\(t\)\\mathbb\{I\}\[t\\in d\]\}\{\\displaystyle\\sum\_\{t\\in\\mathcal\{T\}\_\{q\}^\{\\mathrm\{sp\}\}\}w\_\{\\beta\}\(t\)\}\.\(13\)This score emphasizes discriminative exact matches and complements the capability\-level semantic evidence captured by the dense channel\.

Becauses⟂s\_\{\\perp\}andslexs\_\{\\mathrm\{lex\}\}have different scales, we normalize each channelc∈\{⟂,lex\}c\\in\\\{\\perp,\\mathrm\{lex\}\\\}within the candidate set𝒞q\\mathcal\{C\}\_\{q\}:

s~c​\(q,d\)=sc​\(q,d\)−μc​\(q\)σc​\(q\)\+ϵ,\\widetilde\{s\}\_\{c\}\(q,d\)=\\frac\{s\_\{c\}\(q,d\)\-\\mu\_\{c\}\(q\)\}\{\\sigma\_\{c\}\(q\)\+\\epsilon\},\(14\)whereμc​\(q\)\\mu\_\{c\}\(q\)andσc​\(q\)\\sigma\_\{c\}\(q\)are the corresponding mean and standard deviation over𝒞q\\mathcal\{C\}\_\{q\}\. The final relevance score is

sSkillSight​\(q,d\)=s~⟂​\(q,d\)\+\[s~lex​\(q,d\)\]\+,s\_\{\\mathrm\{SkillSight\}\}\(q,d\)=\\widetilde\{s\}\_\{\\perp\}\(q,d\)\+\\left\[\\widetilde\{s\}\_\{\\mathrm\{lex\}\}\(q,d\)\\right\]\_\{\+\},\(15\)where\[x\]\+=max⁡\(x,0\)\[x\]\_\{\+\}=\\max\(x,0\)\. The background\-calibrated dense score provides the primary ranking signal, while above\-average lexical matches receive a positive correction\. If the lexical scores are constant across𝒞q\\mathcal\{C\}\_\{q\}, the correction becomes zero and ranking relies solely on the dense channel\.

MethodSRA\-BenchSkillBench\-SuppHit@5Recall@10MRR@10LatencyHit@5Recall@10MRR@10LatencyBM2563\.3763\.5151\.543\.4165\.7553\.1557\.1624\.20TF\-IDF Cosine56\.3058\.0442\.562\.3253\.4242\.1146\.2616\.42ColBERTv241\.8842\.3631\.3612\.4256\.1638\.9438\.6931\.54SPLADE55\.2554\.9143\.6040\.5528\.7714\.4022\.8842\.01ABTT67\.7066\.0157\.080\.1278\.0859\.7766\.533\.18SIF35\.4636\.0727\.280\.1139\.7329\.0233\.440\.75Dense67\.7066\.0257\.090\.1375\.3456\.5666\.670\.57BGE\-M3 Hybrid65\.6765\.4655\.2919\.0772\.6058\.6062\.2045\.82Dense–BM25 RRF75\.6777\.1362\.10167\.7473\.9756\.7264\.36219\.26Dense \+ Reranker80\.8777\.1872\.211460\.3576\.7156\.8366\.711740\.79SkillSight86\.0486\.2374\.021\.1776\.7164\.2467\.282\.57Table 1:Main retrieval results on SRA\-Bench and SkillBench\-Supp\. Dense, ABTT, SIF, and SkillSight use Qwen3\-Embedding\-0\.6B, while the reranker uses Qwen3\-Reranker\-0\.6B\. Latency is measured in milliseconds \(ms\)\. The best result for each retrieval metric is highlighted inbold\.

## Experiments

### Experimental Setup

#### Datasets

We conduct retrieval experiments on SRA\-Bench\(Suet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib28)\)and SkillBench\-Supp\(Zhenget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib23)\), both of which provide queries, candidate skill documents, and corresponding gold skills\. SRA\-Bench contains 5,400 test instances and approximately 26k candidate skills, whereas SkillBench\-Supp contains approximately 77k candidate skills, covering retrieval settings at different corpus scales\. We additionally report comparative retrieval results on SkillRet\(Choet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib7)\)in Appendix C of the supplementary material\. End\-to\-end agent evaluation is conducted on SRA\-Bench, whose unified execution protocol enables controlled comparison across retrieval methods\.

#### Baselines

We compare SkillSight with representative retrieval methods covering complementary relevance signals\. BM25\(Robertson and Zaragoza[2009](https://arxiv.org/html/2607.18785#bib.bib22)\)and SPLADE\(Formalet al\.[2021](https://arxiv.org/html/2607.18785#bib.bib21)\)represent sparse retrieval, Dense\(Zhanget al\.[2025](https://arxiv.org/html/2607.18785#bib.bib16)\)represents dense retrieval, and ColBERTv2\(Santhanamet al\.[2022](https://arxiv.org/html/2607.18785#bib.bib20)\)models fine\-grained token interactions\. We further include ABTT\(Mu and Viswanath[2018](https://arxiv.org/html/2607.18785#bib.bib8)\)for representation\-space correction, BGE\-M3 Hybrid\(Chenet al\.[2024a](https://arxiv.org/html/2607.18785#bib.bib15)\)for dense–sparse fusion, and Dense \+ Reranker for two\-stage ranking\. For end\-to\-end evaluation, we compare against LLM Selection and Progressive Disclosure, two representative skill\-incorporation strategies evaluated in SRA\-Bench\(Suet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib28)\)\. LLM Selection uses the agent model to select a skill from the retrieved candidates before execution, whereas Progressive Disclosure allows the agent to load full skill content on demand\. Together, they represent model\-based preselection and dynamic agent\-controlled skill access\.

#### Evaluation Metrics

For offline retrieval evaluation, we useHit​@​k\\mathrm\{Hit\}@k,Recall​@​k\\mathrm\{Recall\}@k, andMRR​@​k\\mathrm\{MRR\}@kto provide a comprehensive assessment of retrieval performance\. These metrics measure whether the top\-kkresults contain a correct skill, the proportion of gold skills covered by the top\-kkresults, and the reciprocal rank of the first correct result, respectively\. Detailed definitions and explanations of the metrics are provided in Appendix B of the supplementary material\.

For end\-to\-end evaluation, we measure agent execution performance according to the task\-specific evaluation protocol of each benchmark\. All methods use the same agent model and execution environment, with only the skill retrieval or selection module varied\.

#### Implementation Details

For each query, SBC constructs an intermediate candidate pool𝒞q\\mathcal\{C\}\_\{q\}of sizeKc=300K\_\{c\}=300, within which semantic and lexical evidence are fused for final ranking\. We setβ=1\\beta=1and determinerrautomatically from the spectral effective rank of the background matrix without evaluation labels\. For end\-to\-end evaluation, the top three retrieved skill documents are inserted into the context of Llama\-3\.1\-8B\-Instruct\(Grattafioriet al\.[2024](https://arxiv.org/html/2607.18785#bib.bib5)\), GPT\-5\.4\-mini, or Qwen3\-4B\-Instruct\(Yanget al\.[2025](https://arxiv.org/html/2607.18785#bib.bib4)\)\. Appendices A and D provide the algorithmic details and hyperparameter analyses, respectively\.

ModelMethodTheoremQALogicBenchToolQACHAMPMedCalcBigCodeBenchOverallLatency \(s\)TokenLlama\-3\.1\-8BLLM Direct31\.0654\.473\.2922\.4227\.5530\.0925\.7210\.424,496Oracle Skill46\.1868\.9516\.8539\.4660\.8236\.1442\.2013\.056,617LLM Selection35\.3454\.874\.6927\.8056\.3634\.8233\.8321\.658,525Progressive Disclosure38\.6948\.420\.3526\.4658\.8231\.1431\.9174\.1810,942SkillSight42\.7061\.0514\.9030\.9458\.6433\.7738\.8019\.917,703GPT\-5\.4\-miniLLM Direct78\.3176\.8441\.0577\.5877\.3649\.1261\.855\.514,581Oracle Skill81\.9392\.1155\.2481\.1790\.6461\.4073\.706\.036,521LLM Selection79\.9280\.1348\.2579\.3786\.8255\.7067\.8317\.0210,591Progressive Disclosure79\.9273\.5544\.2079\.8289\.4553\.4265\.917\.8710,258SkillSight81\.3984\.6145\.6682\.0689\.8260\.2669\.676\.129,205Qwen3\-4BLLM Direct51\.2776\.3227\.4168\.6136\.1842\.4144\.2513\.583,279Oracle Skill67\.4788\.6847\.6271\.7582\.9149\.3064\.6916\.865,985LLM Selection64\.6677\.6335\.5969\.9677\.1842\.3756\.8526\.858,829Progressive Disclosure66\.6779\.3429\.2370\.8567\.7342\.3253\.7930\.859,446SkillSight68\.4080\.5341\.0573\.9980\.9147\.0261\.1320\.518,234

Table 2:End\-to\-end performance on SRA\-Bench across six tasks and three agent models\. Oracle Skill directly provides the annotated gold skills as an reference\. From the top\-50 BM25 candidates, LLM Selection loads one skill before execution, whereas Progressive Disclosure loads skills on demand during inference\. Overall is averaged over all instances, and Token denotes the average number of input tokens per instance\. The best non\-oracle results are bolded\.

### Main Results

#### Offline Retrieval

As shown in Table[1](https://arxiv.org/html/2607.18785#Sx4.T1), SkillSight consistently improves retrieval performance on both datasets, with the largest gains observed inRecall​@​10\\mathrm\{Recall\}@10\. Compared with the original Dense retriever, SkillSight increasesRecall​@​10\\mathrm\{Recall\}@10from 66\.02 to 86\.23 on SRA\-Bench and from 56\.56 to 64\.24 on SkillBench\-Supp\. These results indicate that SkillSight reduces cases in which gold skills are obscured by candidates with similar descriptions\. SkillSight improves all metrics on SRA\-Bench, whereas its gains on SkillBench\-Supp are concentrated primarily in recall\. This difference suggests that its main effect is to increase the coverage of gold skills among the top\-ranked candidates, while improvements at earlier ranks depend more strongly on dataset characteristics\. Overall, the consistent recall gains across datasets demonstrate the complementary effects of background semantic calibration and fine\-grained lexical evidence\. Further retrieval results across different cutoffs and datasets are provided in Appendix C of the supplementary material\.

#### End\-to\-end Agent Execution

As shown in Table[2](https://arxiv.org/html/2607.18785#Sx5.T2), SkillSight achieves the best non\-oracle Overall score across all three models, outperforming LLM Selection by 4\.97, 1\.84, and 4\.28 percentage points, respectively\. The improvement is particularly pronounced on ToolQA with weaker models, reaching 10\.21 percentage points for Llama\-3\.1\-8B and 5\.46 percentage points for Qwen3\-4B\. This result indicates that when candidate skills share similar functional descriptions but differ in specific operations and invocation constraints, SkillSight identifies the target skill more accurately by suppressing bias from shared descriptions and strengthening fine\-grained matching signals\. SkillSight narrows the overall gap to Oracle Skill, showing that more accurate retrieval brings agent performance closer to that achieved when gold skills are provided directly\.

### Efficiency Analysis

SkillSight adds limited overhead to standard dense retrieval\. The background subspaceBBand projected document representationsd⟂d\_\{\\perp\}are precomputed offline\. At inference time, query projection costs𝒪​\(d​r\)\\mathcal\{O\}\(dr\), while lexical matching overCqC\_\{q\}costs𝒪​\(\|Cq\|​\|Tqsp\|\)\\mathcal\{O\}\(\|C\_\{q\}\|\|T\_\{q\}^\{\\mathrm\{sp\}\}\|\)\. The corpus\-wide semantic search retains the standard dense retrieval cost of𝒪​\(N​d\)\\mathcal\{O\}\(Nd\)\. Therefore, SkillSight introduces only𝒪​\(d​r\+\|Cq\|​\|Tqsp\|\)\\mathcal\{O\}\\\!\\left\(dr\+\|C\_\{q\}\|\|T\_\{q\}^\{\\mathrm\{sp\}\}\|\\right\)additional online complexity, without extra neural inference beyond the base encoder or trainable parameters\.

As shown in Table[1](https://arxiv.org/html/2607.18785#Sx4.T1), SkillSight achieves retrieval latencies of only 1\.17 and 2\.57 on the two datasets, respectively\. It is 85–143×\\timesfaster than Dense–BM25 RRF and 677–1,248×\\timesfaster than Dense \+ Reranker, while also achieving better retrieval performance\. In the end\-to\-end setting, Table[2](https://arxiv.org/html/2607.18785#Sx5.T2)shows that SkillSight reduces latency by 8\.0%–64\.0% and token consumption by 6\.7%–13\.1% compared with LLM Selection\.

### Comparison with Specialized Skill Retrievers

#### Full\-Pipeline Comparison\.

As shown in Table[3](https://arxiv.org/html/2607.18785#Sx5.T3), SkillSight outperforms SkillRouter\(Zhenget al\.[2026](https://arxiv.org/html/2607.18785#bib.bib23)\)by 15\.84 Recall@10 points and is1,397×1\{,\}397\\timesfaster on SRA\-Bench\. On SkillBench\-Supp, SkillRouter leads by 2\.13 points, whereas SkillSight is716×716\\timesfaster, demonstrating a favorable accuracy–efficiency trade\-off\. Since the pretrained SkillRet\(Choet al\.[2026](https://arxiv.org/html/2607.18785#bib.bib7)\)reranker checkpoint was unavailable, we evaluate only its released embedding model\.

#### Generalization across Embedding Models\.

Table[4](https://arxiv.org/html/2607.18785#Sx5.T4)shows that SkillSight consistently improves Recall@10 across Qwen, SkillRouter, and SkillRet embeddings, with gains of 8\.21–21\.80 points on SRA\-Bench and 2\.62–7\.68 points on SkillBench\-Supp\. This demonstrates that SkillSight complements retrievers of varying strengths and is not tied to a particular embedding model\.

SRA\-BenchSkillBench\-SuppMethodR@10LatencyR@10LatencySkillRouter70\.391633\.9766\.371839\.97SkillSight86\.231\.1764\.242\.57Table 3:Recall@10 and retrieval latency \(ms\) of SkillSight and the complete SkillRouter retrieval pipeline\. SkillRouter uses its specialized embedding and reranking models, whereas SkillSight requires no neural reranking\.SRA\-BenchSkillBench\-SuppModelDenseSkillSightDenseSkillSightQwen366\.0286\.2356\.5664\.24SkillRouter58\.0179\.8162\.3968\.85SkillRet78\.2186\.4261\.7864\.40Table 4:Recall@10 of Dense and SkillSight with Qwen3\-Embedding\-0\.6B, SkillRouter\-Embedding\-0\.6B, and SkillRet\-Embedding\-0\.6B on SRA\-Bench and SkillBench\-Supp\.MethodHit@5R@10MRR@10Dense retrievalDense67\.7066\.0257\.09Dense \+ LEC81\.5079\.7669\.53Sparse retrievalBM2563\.3763\.5151\.54BM25 \+ SBC72\.2675\.4158\.62Ablation studyw/o LEC \(SBC only\)78\.6178\.2365\.77w/o SBC \(LEC only\)69\.4371\.4654\.28SkillSight86\.0486\.2374\.02Table 5:Component\-level ablation results on SRA\-Bench\. Dense \+ LEC and BM25 \+ SBC evaluate the incremental contribution of lexical and semantic calibration under dense and sparse retrieval paradigms, respectively\. The remaining variants remove the corresponding component from the complete SkillSight framework\.StrategyHit@5R@10MRR@10All82\.1285\.5471\.39Reverse80\.9479\.4868\.28Random84\.7285\.2072\.38SkillSight86\.0486\.2374\.02Table 6:Retrieval performance of different generic token selection strategies on SRA\-Bench\. All, Reverse, and Random estimate the background structure using all tokens, high\-IDF tokens, and randomly selected tokens, respectively, while SkillSight uses low\-IDF generic tokens\.![Refer to caption](https://arxiv.org/html/2607.18785v1/x4.png)Figure 4:End\-to\-end accuracy–latency trade\-off of different methods on SRA\-Bench using Qwen3\-4B\-Instruct\. Circles denote SkillSight with different Top\-kksettings, while squares denote the baselines\. The star marks the default Top\-33setting\. SkillSight uses Top\-33in the main experiments\.

### Ablation Study

#### Effects of the Calibration Components

Table[5](https://arxiv.org/html/2607.18785#Sx5.T5)shows that both components provide effective signals beyond their respective retrieval paradigms\. Adding LEC to Dense retrieval increasesRecall​@​10\\mathrm\{Recall\}@10from 66\.02 to 79\.76, demonstrating that exact matches on discriminative tokens recover fine\-grained capability differences overlooked by dense representations\. Adding SBC to BM25 yields an 11\.90\-point gain inRecall​@​10\\mathrm\{Recall\}@10, indicating that calibrated semantic signals alleviate the sensitivity of lexical retrieval to surface\-form variation\.

Removing LEC from SkillSight reducesMRR​@​10\\mathrm\{MRR\}@10from 74\.02 to 65\.77, whereas removing SBC further decreases it to 54\.28\. This larger degradation identifies shared\-description alignment as a primary source of dense ranking bias, while LEC mainly improves the ordering of semantically similar candidates\. The full model performs best across all metrics, confirming that SBC suppresses corpus\-shared signals and LEC restores task\-specific matching evidence\.

#### Generic Token Selection

As shown in Table[6](https://arxiv.org/html/2607.18785#Sx5.T6), low\-IDF generic tokens achieve the best performance across all metrics\. In contrast, the high\-IDF Reverse strategy reducesRecall​@​10\\mathrm\{Recall\}@10from 86\.23 to 79\.48, indicating that task\-specific tokens are less suitable for background estimation\. Although the All and Random strategies retain strong recall, their lowerHit​@​5\\mathrm\{Hit\}@5andMRR​@​10\\mathrm\{MRR\}@10suggest that generic\-token selection provides a cleaner estimate of the shared background and improves top\-ranked results beyond low\-rank projection alone\.

#### Effect of Top\-kk

Figure[4](https://arxiv.org/html/2607.18785#Sx5.F4)shows how the number of retrieved skills supplied to the agent affects end\-to\-end accuracy and latency\. Increasingkkfrom11to33substantially improves accuracy with limited additional latency, whereas largerkkvalues increase latency and reduce accuracy by introducing functionally similar or irrelevant skills\. Top\-11provides insufficient target coverage, while Top\-33best balances coverage, candidate interference, and inference cost, and is therefore used in the main experiments\.

## Conclusion

In this work, we investigated how shared descriptive patterns in skill documents affect retrieval\. Our analysis reveals that these patterns form a document\-side background that contributes to dense relevance scores, creates an energy gap between queries and skill documents, and obscures task\-relevant signals\. Motivated by this finding, we proposed SkillSight, a training\-free framework that calibrates the shared background in both semantic and lexical spaces\. Experiments on SRA\-Bench and SkillBench\-Supp demonstrate consistent improvements over strong retrieval baselines, while end\-to\-end evaluations confirm that these gains translate into more effective skill invocation with low inference overhead\. Overall, our findings highlight the importance of explicitly accounting for the distinctive structure of skill documents in retrieval\. Future work may extend background calibration to dynamically evolving skill libraries and retrieval models with jointly learned background representations\.

## References

- A simple but tough\-to\-beat baseline for sentence embeddings\.InInternational conference on learning representations,Cited by:[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1)\.
- S\. Bruch, S\. Gai, and A\. Ingber \(2023\)An analysis of fusion functions for hybrid retrieval\.ACM Transactions on Information Systems42\(1\),pp\. 1–35\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1)\.
- J\. Chen, S\. Xiao, P\. Zhang, K\. Luo, D\. Lian, and Z\. Liu \(2024a\)M3\-embedding: multi\-linguality, multi\-functionality, multi\-granularity text embeddings through self\-knowledge distillation\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 2318–2335\.External Links:[Link](https://aclanthology.org/2024.findings-acl.137/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.137)Cited by:[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- Y\. Chen, J\. Yoon, D\. S\. Sachan, Q\. Wang, V\. Cohen\-Addad, M\. Bateni, C\. Lee, and T\. Pfister \(2024b\)Re\-invoke: tool invocation rewriting for zero\-shot tool retrieval\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 4705–4726\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- H\. Cho, R\. Kang, and Y\. Kim \(2026\)SkillRet: a large\-scale benchmark for skill retrieval in llm agents\.arXiv preprint arXiv:2605\.05726\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1),[Datasets](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px1.p1.1),[Full\-Pipeline Comparison\.](https://arxiv.org/html/2607.18785#Sx5.SSx4.SSS0.Px1.p1.2)\.
- R\. Ding, W\. Cheng, M\. Shao, and C\. Zhao \(2026\)SkillGen: learning domain skills for in\-context sequential decision making\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 30512–30520\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p1.1)\.
- T\. Formal, C\. Lassance, B\. Piwowarski, and S\. Clinchant \(2021\)SPLADE v2: sparse lexical and expansion model for information retrieval\.arXiv preprint arXiv:2109\.10086\.Cited by:[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.\(2024\)The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[Implementation Details](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px4.p1.4)\.
- V\. Karpukhin, B\. Oguz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih \(2020\)Dense passage retrieval for open\-domain question answering\.InProceedings of the 2020 conference on empirical methods in natural language processing \(EMNLP\),pp\. 6769–6781\.Cited by:[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1)\.
- X\. Li, Y\. Liu, W\. Chen, B\. You, Z\. Di, Y\. He, S\. Zheng, K\. W\. Choe, J\. Sun, S\. Wang,et al\.\(2026\)SkillsBench: benchmarking how well agent skills work across diverse tasks\.arXiv preprint arXiv:2602\.12670\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- Q\. Liang, H\. Wang, Z\. Liang, and Y\. Liu \(2026\)From skill text to skill structure: the scheduling\-structural\-logical representation for agent skills\.arXiv preprint arXiv:2604\.24026\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1)\.
- D\. Liu, Z\. Li, H\. Du, X\. Wu, S\. Gui, Y\. Kuang, and L\. Sun \(2026\)Graph\-of\-skills: dependency\-aware structural retrieval for massive agent skills\.arXiv preprint arXiv:2604\.05333\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p2.1)\.
- J\. Mu and P\. Viswanath \(2018\)All\-but\-the\-top: simple and effective postprocessing for word representations\.InInternational Conference on Learning Representations,Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1),[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- N\. Otsu \(1979\)A threshold selection method from gray\-level histograms\.IEEE Transactions on Systems, Man, and Cybernetics9\(1\),pp\. 62–66\.External Links:[Document](https://dx.doi.org/10.1109/TSMC.1979.4310076)Cited by:[Query–Document Background Energy Gap](https://arxiv.org/html/2607.18785#Sx3.SSx2.p1.6)\.
- Y\. Qin, S\. Liang, Y\. Ye, K\. Zhu, L\. Yan, Y\. Lu, Y\. Lin, X\. Cong, X\. Tang, B\. Qian, S\. Zhao, L\. Hong, R\. Tian, R\. Xie, J\. Zhou, M\. Gerstein, d\. li, Z\. Liu, and M\. Sun \(2024\)ToolLLM: facilitating large language models to master 16000\+ real\-world apis\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 9695–9717\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/28e50ee5b72e90b50e7196fde8ea260e-Paper-Conference.pdf)Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p1.1),[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- C\. Qu, S\. Dai, X\. Wei, H\. Cai, S\. Wang, D\. Yin, J\. Xu, and J\. Wen \(2024\)Towards completeness\-oriented tool retrieval for large language models\.InProceedings of the 33rd ACM International Conference on Information and Knowledge Management,pp\. 1930–1940\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- S\. Robertson and H\. Zaragoza \(2009\)The probabilistic relevance framework: bm25 and beyond\.Vol\.4,Now Publishers Inc\.Cited by:[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- K\. Santhanam, O\. Khattab, J\. Saad\-Falcon, C\. Potts, and M\. Zaharia \(2022\)Colbertv2: effective and efficient retrieval via lightweight late interaction\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 3715–3734\.Cited by:[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- Z\. Shi, Y\. Wang, L\. Yan, P\. Ren, S\. Wang, D\. Yin, and Z\. Ren \(2025\)Retrieval models aren’t tool\-savvy: benchmarking tool retrieval for large language models\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 24497–24524\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- W\. Su, J\. Long, Q\. Ai, Q\. He, Y\. Tang, C\. Wang, Y\. Tu, Y\. Wang, and Y\. Liu \(2026\)Skill retrieval augmentation for agentic ai\.arXiv preprint arXiv:2604\.24594\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1),[Datasets](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px1.p1.1),[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- N\. Thakur, N\. Reimers, A\. Rücklé, A\. Srivastava, and I\. Gurevych \(2021\)BEIR: a heterogeneous benchmark for zero\-shot evaluation of information retrieval models\.InThirty\-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track \(Round 2\),Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p3.1),[Shared Descriptive Background](https://arxiv.org/html/2607.18785#Sx3.SSx1.p2.1)\.
- P\. Xu, G\. Wu, X\. Chen, T\. Yu, C\. Xiao, F\. Dernoncourt, T\. Zhou, W\. Ai, and V\. Swaminathan \(2026\)Skill discovery for software scripting automation via offline simulations with LLMs\.InFindings of the Association for Computational Linguistics: EACL 2026,V\. Demberg, K\. Inui, and L\. Marquez \(Eds\.\),Rabat, Morocco,pp\. 743–759\.External Links:[Link](https://aclanthology.org/2026.findings-eacl.37/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-eacl.37),ISBN 979\-8\-89176\-386\-9Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[Implementation Details](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px4.p1.4)\.
- S\. Yuan, K\. Song, J\. Chen, X\. Tan, Y\. Shen, K\. Ren, D\. Li, and D\. Yang \(2025\)Easytool: enhancing llm\-based agents with concise tool instruction\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 951–972\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p1.1)\.
- K\. Zeng, Y\. Huo, S\. Zhang, Z\. Ye, Y\. Zhuo, H\. Liu, Y\. Lu, J\. Wen, and X\. Tang \(2026\)Group of skills: group\-structured skill retrieval for agent skill libraries\.arXiv preprint arXiv:2605\.06978\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p2.1)\.
- Y\. Zhang, M\. Li, D\. Long, X\. Zhang, H\. Lin, B\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Lin,et al\.\(2025\)Qwen3 embedding: advancing text embedding and reranking through foundation models\.arXiv preprint arXiv:2506\.05176\.Cited by:[Baselines](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px2.p1.1)\.
- Y\. Zheng, Z\. Zhang, C\. Ma, Y\. Yu, J\. Zhu, Y\. Wu, T\. Xu, B\. Dong, H\. Zhu, R\. Huang,et al\.\(2026\)Skillrouter: skill routing for llm agents at scale\.arXiv preprint arXiv:2603\.22455\.Cited by:[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1),[Datasets](https://arxiv.org/html/2607.18785#Sx5.SSx1.SSS0.Px1.p1.1),[Full\-Pipeline Comparison\.](https://arxiv.org/html/2607.18785#Sx5.SSx4.SSS0.Px1.p1.2)\.
- Y\. Zheng, P\. Li, W\. Liu, Y\. Liu, J\. Luan, and B\. Wang \(2024\)Toolrerank: adaptive and hierarchy\-aware reranking for tool retrieval\.InProceedings of the 2024 joint international conference on computational linguistics, language resources and evaluation \(LREC\-COLING 2024\),pp\. 16263–16273\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Agent Tool and Skill Retrieval](https://arxiv.org/html/2607.18785#Sx2.SSx1.p1.1)\.
- X\. Zhou, Y\. Gao, X\. Jie, X\. Cai, J\. Bu, and H\. Wang \(2024\)Ease\-dr: enhanced sentence embeddings for dense retrieval\.InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval,pp\. 2374–2378\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1)\.
- S\. Zhuang, X\. Ma, B\. Koopman, J\. Lin, and G\. Zuccon \(2024\)PromptReps: prompting large language models to generate dense and sparse representations for zero\-shot document retrieval\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 4375–4391\.Cited by:[Introduction](https://arxiv.org/html/2607.18785#Sx1.p2.1),[Retrieval Representations and Relevance Calibration](https://arxiv.org/html/2607.18785#Sx2.SSx2.p1.1)\.