Research map / Large language models / Language, safety and interpretability

Language Ranker: A Lightweight Ranking framework for LLM Decoding

NeurIPS 2025 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Language Ranker, a lightweight (<0.5M parameter) listwise or pointwise ranker that reranks LLM-sampled candidates using hidden-state features from the base model, matching reward models on MATH, MBPP, and xLAM function calling.

Problem

Decoding of LLM responses is largely rule-based, and reward-model reranking of sampled candidates is computationally expensive in training and inference.

Approach

Treats decoding as the ranking stage of a recommender pipeline: the base model's intermediate hidden states at the final token of the instruction and each candidate serve as shared features, and a tiny Transformer (listwise) or MLP (pointwise) ranker reranks the candidates, so no extra reward-model forward pass is needed.

Key results

• On Llama3.1-8B-Instruct, ListRanker reaches 46.3 on MATH and 54.5 on MBPP versus 25.1 and 41.9 for first sample and 45.1 and 52.9 for the Llama8B reward model.
• On Qwen2.5-32B-Instruct, rankers with under 0.5M parameters reach 81.1 on MATH, comparable to the 32.8B RM at 80.7.
• Rankers trained on MATH or MBPP transfer across tasks, with the MATH-trained ranker scoring 43.4 on MBPP vs 47.7 for the gpt2 RM.

Measured effects reported in the paper

Surprising findings

• A ranker trained on one MATH subject transfers to other subjects with only small drops, sometimes approaching in-domain performance.
• Increasing ranker block count from 1 to 4 yields only marginal gains, and removing the projection layer gives little improvement despite a much larger ranker.

Limitations

• Requires access to final-token hidden states at inference, which is not supported by inference frameworks such as vLLM.
• Gains over reward models are not uniform; on xLAM the Llama8B reward model is slightly better (32.8 vs 32.6), and on MBPP for Qwen32B the RM is best.
• Evaluated only on a few tasks with 1,000-sample training sets and 100 candidates per query.

Experimental setup

Datasets: MATH MBPP xlam-function-calling-60k Databricks-Dolly-15k AlpacaEval

Benchmarks: MATH MBPP xLAM function calling AlpacaEval

Metrics: accuracy length-controlled win rate

Baselines: RM (gpt2) RM (Llama8B) RM (Qwen7B) RM (Qwen32B) Beam Search First Sample Self-Consistency

Builds on: self-consistency reward modeling LoRA

Closest papers by meaning

• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.90)
• Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · similarity 0.89)
• Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.89)
• ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.89)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.89)
• SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.89)