L3TR: Mitigating Position Bias and Lost-in-the-Middle in LLM-Based Talent Recommendations for Enhanced Robustness

L3TR (Listwise LLM Talent Recommendation Framework) targets the position bias and lost-in-the-middle problems in LLM-based recommendations. It proposes block attention mechanism, local position encoding, and ID sampling methods, adopts the listwise paradigm to process candidate lists in one go, and has been validated effective on real-world datasets, improving recommendation fairness and robustness.

Talent recruitment has high costs and long cycles, while traditional HR screening is inefficient. LLMs have great application potential but have flaws: existing pointwise paradigms involve repeated processing, lack comparison, and have fragmented context; more critically, LLMs' position bias (focusing on the beginning and end while ignoring the middle) and lost-in-the-middle (declined ability to extract middle information) lead to systematic undervaluation of middle candidates.

L3TR adopts the listwise paradigm to process the entire candidate list at once for relative comparison and ranking. Its advantages include: understanding candidates' relative strengths and weaknesses from a global perspective, efficient processing to reduce repeated overhead, and natural expression of comparative reasoning; however, it needs to address position bias and lost-in-the-middle issues.

1. Block Attention: Divide the candidate list into blocks, with intra-block interaction + inter-block aggregation, local focus + global integration to alleviate position decay; 2. Local Position Encoding: Encode relative positions only within candidates to avoid absolute position bias; 3. ID Sampling: Randomly sample candidate subsets of different sizes during training to improve generalization ability.

L3TR includes a complete set of tools: position bias detection (score fluctuations of the same candidate in different positions), Token bias detection (analyzing attention distribution based on resume length/complexity), and training-free debiasing (randomly shuffling order and aggregating results to offset bias), which can be applied to other LLM recommendation systems.

Evaluated on two real-world talent datasets (including job positions/resumes, multiple industries, and annotated matching degrees), with metrics including recommendation accuracy, ranking quality, and fairness; results are significantly better than baselines, with improved fairness (enhanced evaluation quality of middle candidates); ablation experiments prove the value of each component (performance decreases when any component is removed).

1. Paradigm Shift: Listwise has great potential in fine-grained comparison scenarios (job/product/content recommendation); 2. General Solution for Position Bias: Block attention and local encoding can be applied to search/advertising/news recommendation; 3. Interpretability: LLM output of comparison reasons improves transparency.

Limitations: High computational overhead, decreased feasibility when candidate set size is large, unvalidated domain adaptability, insufficient handling of dynamic preferences; Future directions: Reduce inference cost, integrate coarse-ranking and fine-ranking architecture, cross-domain validation, and adapt to dynamic needs.