Research map / Large language models / Reasoning and chain of thought

Rethinking Personalization in Large Language Models at the Token Level

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning · 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 PerContrast, a causal-intervention token score (persona-removed vs. full-persona log-prob gap), and PerCE, an EM-style online reweighted cross-entropy that upweights personal tokens for personalized LLM fine-tuning.

Problem

Standard cross-entropy treats all response tokens equally, diluting the training signal for the few tokens that carry user-specific personalization; the token-level personalization degree is not directly observable.

Approach

Estimate each token's personalization degree as the log-probability difference with vs. without the user persona (Personal Influence Ratio, shown to equal a token-level causal effect), clip it, and use it as a per-token weight in an online EM-like weighted CE loss that requires only one extra forward pass.

Key results

• PerCE achieves up to +68.04% METEOR gain over CE on Qwen3-4B for personalized review writing (PRW) on LongLaMP, with average ROUGE-L gain of 15.06% on Qwen3-4B and 10.62% on Llama3-8B.
• On the synthetic personal-token identification task, PerContrast achieves the best precision, recall, and F1 versus word-matching and LLM-as-a-judge baselines; word matching reaches only 26.5% recall.
• Cross-scenario transfer to ALOE improves average LLM-judge score from 1.27 to 2.78 on Qwen3-4B trained on PRW.

Surprising findings

• PerCE slightly underperforms CE on ROUGE-L for Personalized Abstract Generation (-2.90% on Qwen3-4B), while improving METEOR.
• Out-of-domain PerCE scores sometimes exceed in-domain CE scores, e.g., PAG-to-PRW transfer 0.2211 vs in-domain CE 0.1898.

Limitations

• Average-based gains are reported over a sampled subset (2,000 train / 500 test per task) and a small ALOE set of 100 conversations.
• Evaluation relies mainly on word-overlap metrics with an LLM judge only in an appendix; PAG shows a ROUGE-L decrease.
• Weight clipping range [0.8, 5.0] was chosen heuristically and is not optimal for every task.

Experimental setup

Datasets: LongLaMP ALOE LaMP HotpotQA DROP Synthetic-Persona-Chat

Benchmarks: LongLaMP ALOE LaMP HotpotQA DROP

Metrics: ROUGE-L METEOR LLM-as-a-judge score (1-5) precision recall F1

Baselines: CE LossCE EntCE

Builds on: LongLaMP LaMP Contriever retrieval-augmented generation

Closest papers by meaning

• HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.89)
• Rethinking Personalized Generation: Test-time Alignment via Factorized Ranking Models (NeurIPS 2026 · similarity 0.89)
• Context Steering: Controllable Personalization at Inference Time (ICLR 2025 · similarity 0.88)
• Logits are All We Need to Adapt Closed Models (ICML 2025 · similarity 0.88)
• Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.87)
• Model-Aware Tokenizer Transfer (NeurIPS 2026 · similarity 0.87)