Towards Position-Robust Talent Recommendation via Large Language Models

Silin Du; Hongyan Liu

cs.CLApr 2, 2026

Towards Position-Robust Talent Recommendation via Large Language Models

Authors: Silin Du, Hongyan Liu

Organizations: Department of Management Science and Engineering, Tsinghua University, Beijing, China.

Abstract

Talent recruitment is a critical, yet costly process for many industries, with high recruitment costs and long hiring cycles. Existing talent recommendation systems increasingly adopt large language models (LLMs) due to their remarkable language understanding capabilities. However, most prior approaches follow a pointwise paradigm, which requires LLMs to repeatedly process some text and fails to capture the relationships among candidates in the list, resulting in higher token consumption and suboptimal recommendations. Besides, LLMs exhibit position bias and the lost-in-the-middle issue when answering multiple-choice questions and processing multiple long documents. To address these issues, we introduce an implicit strategy to utilize LLM's potential output for the recommendation task and propose L3TR, a novel framework for listwise talent recommendation with LLMs. In this framework, we propose a block attention mechanism and a local positional encoding method to enhance inter-document processing and mitigate the position bias and concurrent token bias issue. We also introduce an ID sampling method for resolving the inconsistency between candidate set sizes in the training phase and the inference phase. We design evaluation methods to detect position bias and token bias and training-free debiasing methods. Extensive experiments on two real-world datasets validated the effectiveness of L3TR, showing consistent improvements over existing baselines.

Figures & tables

Figure 1. The Prompt Template for Listwise Talent Recommendation.

Panel of Figure 3 — Figure 3. Attention Map of Block Attention and Its Variant.

Figure 4. Illustration of Global and Local Position Information.

Figure 5. Architecture of Our Proposed LLM-based Recommendation Model.

Figure 6. Performance with Different

Q

Dataset	# Resumes	# Tokens in Resume	# Jobs	# Tokens in Job Posting	# Records
HRT	29057	1214	1198	352	30810
JobRec	4821	44	10642	402	14868

Table 1. Dataset Statistics.

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.3430	.4236	.1551	.5248	.3348
	BPJFNN	.5077	.5645	.2607	.7409	.4614
	APJFNN	.3872	.4857	.1683	.6056	.3646
	TRBERT	.6078	.6604	.3927	.7937	.5716
LLM-based Models	List-Imp (6B)	.2087	.3062	.0792	.3564	.2276
	List-Exp ( $>$ 100B)	.7982	.8166	.6815	.8311	.7781

Table 2. Recommendation Accuracy on HRT.

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.2080	.3132	.0704	.3527	.2285
	BPJFNN	.2008	.3051	.0669	.3406	.2225
	TRBERT	.1967	.2976	.0663	.3351	.2197
LLM-based Models	List-Imp (20B)	.5970	.6444	.4620	.7191	.5860
	List-Exp ( $>$ 100B)	.7539	.7719	.5994	.8844	.7234
	Point-Imp (20B)	.7119	.7466	.5386	.8616	.6790

Table 3. Recommendation Accuracy on JobRec.

Model	Block Attention	Random ID Assignment	Local Positional Encoding	ND@5	ND@10	R@1	R@5	MRR
$\text{L3TR}^{-}$	$\times$	$\times$	$\times$	.7692	.7999	.7046	.8284	.7682
L3TR-B	$\times$	$\surd$	$\surd$	.6248	.6650	.5248	.7063	.6260
L3TR-I	$\surd$	$\times$	$\surd$	.8156	.8501	.7721	.8516	.8195
L3TR-L	$\surd$	$\surd$	$\times$	.8204	.8419	.7640	.8696	.8177
L3TR	$\surd$	$\surd$	$\surd$	.8379	.8545	.7953	.8762	.8377

Table 4. Ablation Results on HRT.

Model	Block Attention	Random ID Assignment	Local Positional Encoding	ND@5	ND@10	R@1	R@5	MRR
$\text{L3TR}^{-}$	$\times$	$\times$	$\times$	.8894	.9022	.8295	.9351	.8812
L3TR-B	$\times$	$\surd$	$\surd$	.8157	.8576	.7889	.8358	.8232
L3TR-I	$\surd$	$\times$	$\surd$	.9023	.9097	.8581	.9402	.9017
L3TR-L	$\surd$	$\surd$	$\times$	.9031	.9125	.8665	.9410	.9065
L3TR	$\surd$	$\surd$	$\surd$	.9124	.9208	.8683	.9638	.9167

Table 5. Ablation Results on JobRec.

Figure 7. Accuracy Fluctuations with Different

Q

Panel of Figure 9 — Figure 9. Token Bias at Different Positions of L3TR and L3TR - .

Panel of Figure 10 — Figure 10. Token Bias of L3TR and L3TR - .

Model	ND@5	ND@10	R@1	R@5	MRR
List-Exp	.7982	.8166	.6815	.8311	.7781
+ Pre-ranker	.7989	.8195	.6997	.8729	.7870
List-Imp	.2087	.3062	.0792	.3564	.2276
+ Pre-ranker	.3946	.4750	.2277	.5660	.3864
+ Debiasing Method	.2165	.3102	.0859	.3616	.2303
+ Pre-ranker and Debiasing Method	.3972	.4788	.2524	.5437	.3975

Table 7. Evaluations of Different Debiasing Methods on HRT.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
Problem formulation and ranking
$\mathcal{G}$	Set of job postings.
$\mathcal{R}$	Set of resumes (talent pool).
$g_{i}$	The $i$ -th job posting, $g_{i}\in\mathcal{G}$ .
$r_{j}$	The $j$ -th resume, $r_{j}\in\mathcal{R}$ .
$\mathcal{M}$	Historical matching records set, $\mathcal{M}=\{(g_{i},r_{j},y_{g_{i},r_{j}})\}$ .

Appendix

Table 8. Notation and Symbol Definitions in the Main Text.

Figure 11. Long-term Decay of RoPE (from Su et al. (2024) ).

Figure 12. The Variation of

\frac{1}{d/4}\sum^{d/2}_{i=d/4+1}|S_{i}^{\prime}|

with Respect to Relative Distance

m^{\prime}-n

(

d=128

Figure 13. The Variation of

\frac{1}{d/2}\sum_{i=0}^{d/2}\left|S_{i}\right|

with Respect to Relative Distance

m^{\prime}-n

(

d=128

Job	Description	Resume	Description
Title	Sales Manager	Education	Bachelor degree in Electronic Science and Technology
Company	A company for enterprise employee training	Current Position	General Sales Manager
Duty	1. Develop business customer cooperation across the country through a variety of channels (telephone, network, exhibition, chamber of commerce, will sell, visit, friend recommendation, etc.); 2. Mine customer needs and develop customers according to the company’s product characteristics; 3. Introduce and promote the company’s mobile learning platform to enterprise customers; 4. Design learning plans and customize new media courses for customers according to their needs; 5. Complete the established customer development and sales targets on time; 6. Complete the work assigned by the leader	Job Track	1. General sales manager of a company for mobile CRM software in Shenzhen from 2015 to 2019; 2. Area channel director of an Internet marketing software company from 2012 to 2015; 3. Co-founder of a technology company for website construction, searching engine optimization software, nesting bidding promotion from 2011 to 2012; 4. Sales manager in a famous website providing efficient network promotion plans from 2009 to 2011; 5. Sales representative in a well-known talent website from 2005 to 2009.
Requirement	1. 3 years and above work experience, master the systematic sales method; 2. With the experience and ability to sell complex solutions for large customers; 3. Strong executive ability, love learning, with strong analysis and summary ability; 4. Have a clear self-cognition and goal direction, can withstand greater work pressure; 5. Traditional IT industry distribution experience is preferred; 6. Can adapt to business trips and other emergencies
Location	Shenzhen, China	Workplace	Shenzhen, China

Appendix

Table 9. A Pair of Matched Resume and Job Posting from HRT.

Job	Description	Resume	Description
Title	Automotive Service Advisor	Degree Type	High School
Description	Briggs Nissan in Lawrence Kansas has an immediate opportunity for the right person to join our team as a Service Advisor. Due to growth and expansion in our department we are looking for the right person that is looking for a career with our team. Outstanding customer relation skills required. ADP computer experience a plus. Service department experience a plus but not required. Some Saturdays required. Excellent benefit package. If you feel you have what it takes be become part of our team. Please contact, Steve Schons, Service Manager Briggs Nissan, Lawrence Ks. (785) 856-7199	Total Work Experience	5.0 years

		Currently Employed	Yes

		In a Management Position	Yes

Appendix

Table 10. A Pair of Matched Resume and Job Posting from JobRec.

You are responsible for recruiting for a vacant position in a company. Please analyze whether the candidate is suitable for this position, based on the candidate’s resume and the job posting.

The job posting is as follows.

\{g\}

Resume

\{c_{1}^{g}\}

Please output in a JSON format. The referenced output is as follows.

{"score": the matching score of the candidate and the job, which should be between 0 and 1.}

Appendix

Table 11. Pointwise Recommendation Template for HRT.

You are responsible for recruiting for a vacant position in a company, please rank the resumes of

\{N\}

candidates based on the job description. The ranking should be determined by the matching degree between the resume and the job posting, with candidates who have a higher matching degree placed at the top.

The job posting is as follows.

\{g\}

Here are the resumes for

\{N\}

candidates, each with a unique ID. Please use the IDs to sort all the resumes.

Resume ID: A

Content:

\{c_{1}^{g}\}

Resume ID: B

Content:

\{c_{2}^{g}\}

Appendix

Table 12. Listwise Recommendation Template for HRT.

You are an expert at recruitment. Please analyze whether the candidate would be interested in this job based on the information of the candidate and the job posting.

The detailed information of the candidate is as follows.

\{\}

The job posting is as follows.

\{\}

Please output in a JSON format. The referenced output is as follows.

{"score": the matching score of the candidate and the job, which should be between 0 and 1.}

Appendix

Table 13. Pointwise Recommendation Template for JobRec.

You are an expert at recruitment. Please rank the following

\{N\}

job postings based on the information of the candidate. Please list job postings that align more closely with the candidate’s interests at the front of the list.

The detailed information of the candidate is as follows.

\{\}

The following are

\{N\}

job postings, each with an ID, which is used to rank all jobs.

Job Posting ID: A

Content:

\{\}

Job Posting ID: B

Content:

\{\}

Appendix

Table 14. Listwise Recommendation Template for JobRec.

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.3512	.4319	.1613	.5336	.3421
	BPJFNN	.5189	.5754	.2684	.7521	.4723
	APJFNN	.3965	.4956	.1759	.6148	.3728
	TRBERT	.6186	.6713	.4015	.8052	.5839
LLM-based Models	List-Imp	.2196	.3184	.0857	.3698	.2389
	List-Exp	.8169	.8335	.7018	.8476	.7964

Appendix

Table 15. Recommendation Accuracy on HRT with

N=10

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.3316	.4119	.1478	.5112	.3247
	BPJFNN	.4959	.5512	.2483	.7214	.4489
	APJFNN	.3749	.4723	.1589	.5881	.3526
	TRBERT	.5917	.6438	.3784	.7716	.5534
LLM-based Models	List-Imp	.1958	.2876	.0713	.3349	.2128
	List-Exp	.7716	.7924	.6519	.8047	.7485

Appendix

Table 16. Recommendation Accuracy on HRT with

N=20

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.3116	.3925	.1353	.4879	.3058
	BPJFNN	.4689	.5247	.2328	.6894	.4202
	APJFNN	.3538	.4519	.1471	.5606	.3314
	TRBERT	.5629	.6158	.3576	.7362	.5221
LLM-based Models	List-Imp	.1698	.2534	.0601	.3011	.1849
	List-Exp	.7234	.7471	.6017	.7623	.6992

Appendix

Table 17. Recommendation Accuracy on HRT with

N=30

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.2143	.3208	.0786	.3819	.2387
	BPJFNN	.2071	.3134	.0745	.3698	.2321
	TRBERT	.2024	.3056	.0739	.3652	.2294
LLM-based Models	List-Imp	.6184	.6639	.5037	.7485	.6073
	List-Exp	.7812	.7964	.6521	.9126	.7489
	Point-Imp	.7368	.7689	.5914	.8897	.7058

Appendix

Table 18. Recommendation Accuracy on JobRec with

N=10

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.2049	.3094	.0676	.3474	.2244
	BPJFNN	.1978	.3014	.0642	.3355	.2185
	TRBERT	.1937	.2940	.0636	.3301	.2157
LLM-based Models	List-Imp	.5880	.6367	.4435	.7083	.5755
	List-Exp	.7426	.7626	.5754	.8711	.7104
	Point-Imp	.7012	.7376	.5171	.8487	.6668

Appendix

Table 19. Recommendation Accuracy on JobRec with

N=20

Type	Model	ND@5	ND@10	R@1	R@5	MRR
Domain-specific Models	PJFNN	.2006	.3031	.0638	.3392	.2187
	BPJFNN	.1939	.2956	.0604	.3281	.2129
	TRBERT	.1896	.2887	.0597	.3215	.2103
LLM-based Models	List-Imp	.5732	.6234	.4159	.6861	.5587
	List-Exp	.7214	.7438	.5362	.8475	.6899
	Point-Imp	.6801	.7176	.4853	.8224	.6445

Appendix

Table 20. Recommendation Accuracy on JobRec with

N=30

Model	$N_{\mathrm{train}}$	ND@5	ND@10	R@1	R@5	MRR
L3TR -	3	.7671	.8010	.6960	.8290	.7629
	4	.7692	.7999	.7046	.8284	.7682
	5	.7640	.7965	.6955	.8230	.7625
	6	.7585	.7920	.6900	.8190	.7570
L3TR	3	.8235	.8490	.7780	.8695	.8250
	4	.8379	.8545	.7953	.8762	.8377

Appendix

Table 21. Recommendation Accuracy on HRT with Different

N_{\mathrm{train}}

Model	$S$	ND@5	ND@10	R@1	R@5	MRR
L3TR	0	.8143	.8366	.7706	.8548	.8159
	1	.8185	.8375	.7706	.8614	.8211
	5	.8262	.8473	.7872	.8630	.8257
	10	.8379	.8545	.7953	.8762	.8377
	15	.8291	.8512	.7859	.8688	.8310

Appendix

Table 22. Recommendation Accuracy on HRT with Different

S

ID Token	List-Imp	List-Imp+Pre-ranker	L3TR -	L3TR - +Pre-ranker	L3TR	L3TR+Pre-ranker
“A”	0.4739	0.5168	0.0826	0.3291	0.0563	0.3133
“B”	0.0424	0.0378	0.0915	0.1278	0.0842	0.1357
“C”	0.0309	0.0318	0.1451	0.1396	0.0770	0.1011
“D”	0.2786	0.2426	0.0867	0.0681	0.0636	0.0602
“E”	0.0204	0.0201	0.1120	0.0667	0.0829	0.0579
“F”	0.0243	0.0267	0.0241	0.0150	0.0654	0.0514

Appendix

Table 23. Estimated Prior Bias on HRT with

N=15

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 508 papers

Agentic AI for Human Resources: LLM-Driven Candidate Assessment
Mar 17, 2026Kamer Ali Yuksel, Abdul Basit Anees, Ashraf Elneima +3Learning to Rank LLM-as-a-Judge
aiXplain, Inc., San Jose, CA, USA
Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences)
Mar 18, 2026Zichen Tang, Zirui Zhang, Qian Wang +3Large Language Model-Based Recommendation LLM Agent Reliability
The Hong Kong University of Science and Technology · National University of Singapore · The Hong Kong University of Science and Technology (Guangzhou)
LBR: Towards Mitigating Length Bias in Large Language Models for Recommendation
Jul 5, 2026Hongchen Li, Bohao Wang, Jingbang Chen +5Large Language Model-Based Recommendation Attention Mechanisms
State Key Laboratory of Blockchain and Data Security, Zhejiang University. · College of Computer Science and Technology, Zhejiang University. · Zhejiang University Hangzhou, Zhejiang, China +3

Mar 17, 2026J/K move · Enter open · S save

Agentic AI for Human Resources: LLM-Driven Candidate Assessment

Kamer Ali Yuksel, Abdul Basit Anees, Ashraf Elneima, Sanjika Hewavitharana+2

aiXplain, Inc., San Jose, CA, USA

Learning to Rank LLM-as-a-Judge

Open paper AskPDF arXiv

In this work, we present a modular and interpretable framework that uses Large Language Models (LLMs) to automate candidate assessment in recruitment. The system integrates diverse sources, including job descriptions, CVs, interview transcripts, and HR feedback; to generate structured evaluation reports that mirror expert judgment. Unlike traditional ATS tools that rely on keyword matching or shallow scoring, our approach employs role-specific, LLM-generated rubrics and a multi-agent architecture to perform fine-grained, criteria-driven evaluations. The framework outputs detailed assessment reports, candidate comparisons, and ranked recommendations that are transparent, auditable, and suitable for real-world hiring workflows. Beyond rubric-based analysis, we introduce an LLM-Driven Active Listwise Tournament mechanism for candidate ranking. Instead of noisy pairwise comparisons or inconsistent independent scoring, the LLM ranks small candidate subsets (mini-tournaments), and these listwise permutations are aggregated using a Plackett-Luce model. An active-learning loop selects the most informative subsets, producing globally coherent and sample-efficient rankings. This adaptation of listwise LLM preference modeling (previously explored in financial asset ranking) provides a principled and highly interpretable methodology for large-scale candidate ranking in talent acquisition.

Abstract

Recommender Systems

Explore similar work

Agentic AI for Human Resources: LLM-Driven Candidate Assessment

Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences)

LBR: Towards Mitigating Length Bias in Large Language Models for Recommendation