Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning
read the original abstract
Token-level reweighting is a simple yet effective mechanism for controlling supervised fine-tuning, but common indicators are largely one-dimensional: the ground-truth probability reflects downstream alignment, while token entropy reflects intrinsic uncertainty induced by the pre-training prior. Ignoring entropy can misidentify noisy or easily replaceable tokens as learning-critical, while ignoring probability fails to reflect target-specific alignment. RankTuner introduces a probability--entropy calibration signal, the Relative Rank Indicator, which compares the rank of the ground-truth token with its expected rank under the prediction distribution. The inverse indicator is used as a token-wise Relative Scale to reweight the fine-tuning objective, focusing updates on truly under-learned tokens without over-penalizing intrinsically uncertain positions. Experiments on multiple backbones show consistent improvements on mathematical reasoning benchmarks, transfer gains on out-of-distribution reasoning, and pre code generation performance over probability-only or entropy-only reweighting baselines.
This paper has not been read by Pith yet.
Forward citations
Cited by 5 Pith papers
-
A3M: Adaptive, Adversarial and Multi-Objective Learning for Strategic Bidding in Repeated Auctions
A3M integrates adaptive DRL, adversarial opponent modeling, and multi-objective rewards to cut regret 30-40% versus baselines while remaining robust to strategy shifts in repeated auctions.
-
PriFT: Prior-Support Guided Supervised Fine-Tuning
PriFT uses token reweighting signals from a frozen pretrained model to stabilize SFT and achieve better results than standard SFT baselines on reasoning tasks.
-
EGAD: Entropy-Guided Adaptive Distillation for Token-Level Knowledge Transfer
EGAD adaptively distills LLM knowledge at the token level by using entropy to create a curriculum from low- to high-entropy tokens, adjust temperature, and switch between logits-only and feature-based branches.
-
EVLA: An Electro-Aware Multimodal Assistant for Physically-Grounded Driving Reasoning and Control
EVLA combines a Unified Co-State Encoder and Electro-aware Structured Reasoning Chain with physics-guided training to produce energy-optimal driving decisions, reporting +5.6% accuracy gains over fine-tuned VLM baseli...
-
FinInvest-GTCN: Explainable Graph-Temporal-Causal Modeling for Risk-Aware Investment Decision Optimization
FinInvest-GTCN combines graph, temporal, and causal networks with meta-causal adaptation to improve risk-adjusted predictions for VC investments, achieving RA-MSE of 2.51 and 18.7% higher simulated returns on propriet...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.