On the Unreasonable Effectiveness of Last-layer Retraining

• URL: http://arxiv.org/abs/2512.01766v1
• Date: Mon, 01 Dec 2025 15:08:43 GMT
• Title: On the Unreasonable Effectiveness of Last-layer Retraining
• Authors: John C. Hill, Tyler LaBonte, Xinchen Zhang, Vidya Muthukumar,
• Abstract summary: Last-layer retraining (LLR) methods have garnered interest as an efficient approach to rectify dependence on spurious correlations.<n>LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set.<n>We show how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.
• Score: 11.989603982988344
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Last-layer retraining (LLR) methods -- wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training -- have garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this ``unreasonable effectiveness'' of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in the implicit bias of gradient descent benefiting robustness. Our empirical investigation does not support this hypothesis. Instead, we present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set. We conclude by showing how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.

Related papers

• From Absolute to Relative: Rethinking Reward Shaping in Group-Based Reinforcement Learning [7.6602542594279335]
  We propose Reinforcement Learning with Relative Rewards to shift reward shaping from absolute scoring to relative ranking.<n>We show that RLRR yields consistent performance improvements over standard group-based baselines across reasoning benchmarks and open-ended generation tasks.
  arXiv  Detail & Related papers  (2026-01-30T15:07:06Z)
• A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization [58.116300485427764]
  Reinforcement learning post-training can elicit reasoning behaviors in large language models.<n> token-level correction often leads to unstable training dynamics when the degree of off-policyness is large.<n>We propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO)
  arXiv  Detail & Related papers  (2026-01-30T08:47:19Z)
• Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering [50.63386303357225]
  We propose AdaRAS, a lightweight test-time framework that improves reasoning reliability by selectively intervening on neuron activations.<n>AdaRAS identifies Reasoning-Critical Neurons (RCNs) via a polarity-aware mean-difference criterion and adaptively steers their activations during inference.<n> Experiments on 10 mathematics and coding benchmarks demonstrate consistent improvements, including over 13% gains on AIME-24 and AIME-25.
  arXiv  Detail & Related papers  (2026-01-27T17:53:01Z)
• Your Group-Relative Advantage Is Biased [74.57406620907797]
  Group-based learning methods rely on group-relative advantage estimation to avoid learned critics.<n>In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage.<n>We propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics.
  arXiv  Detail & Related papers  (2026-01-13T13:03:15Z)
• CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs [53.749193998004166]
  Curriculum learning plays a crucial role in enhancing the training efficiency of large language models.<n>We propose CurES, an efficient training method that accelerates convergence and employs Bayesian posterior estimation to minimize computational overhead.
  arXiv  Detail & Related papers  (2025-10-01T15:41:27Z)
• On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization [52.76330545825083]
  Reinforcement learning (RL) has become popular in enhancing the reasoning capabilities of large language models (LLMs)<n>We identify a previously unrecognized phenomenon we term Lazy Likelihood Displacement (LLD), wherein the likelihood of correct responses marginally increases or even decreases during training.<n>We develop a method called NTHR, which downweights penalties on tokens contributing to the LLD. Unlike prior DPO-based approaches, NTHR takes advantage of GRPO's group-based structure, using correct responses as anchors to identify influential tokens.
  arXiv  Detail & Related papers  (2025-05-24T18:58:51Z)
• Bridging Supervised Learning and Reinforcement Learning in Math Reasoning [55.889740979706815]
  Reinforcement Learning (RL) has played a central role in the recent surge of math abilities by enabling self-improvement through binary verifier signals.<n>In this work, we propose Negative-aware Fine-Tuning (NFT) -- a supervised approach that enables LLMs to reflect on their failures and improve autonomously with no external teachers.
  arXiv  Detail & Related papers  (2025-05-23T17:17:40Z)
• A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce [68.99924691391048]
  We revisit GRPO from a reinforce-like algorithm perspective and analyze its core components.<n>We find that a simple rejection sampling baseline, RAFT, yields competitive performance than GRPO and PPO.<n>Motivated by this insight, we propose Reinforce-Rej, a minimal extension of policy gradient that filters both entirely incorrect and entirely correct samples.
  arXiv  Detail & Related papers  (2025-04-15T16:15:02Z)
• Adaptive Group Policy Optimization: Towards Stable Training and Token-Efficient Reasoning [6.417080496525403]
  We propose Adaptive Group Policy Optimization (AGPO) which uses a simple but effective method, an adaptive loss function, to mitigate training fluctuation and token inefficiency.<n>Experiments demonstrate our method achieves more stable training and superior performance with significantly fewer tokens in reasoning steps.
  arXiv  Detail & Related papers  (2025-03-20T08:48:57Z)
• Annotation-Free Group Robustness via Loss-Based Resampling [3.355491272942994]
  Training neural networks for image classification with empirical risk minimization makes them vulnerable to relying on spurious attributes instead of causal ones for prediction. We propose a new method, called loss-based feature re-weighting (LFR), in which we infer a grouping of the data by evaluating an ERM-pre-trained model on a small left-out split of the training data. For a complete assessment, we evaluate LFR on various versions of Waterbirds and CelebA datasets with different spurious correlations.
  arXiv  Detail & Related papers  (2023-12-08T08:22:02Z)
• Unbiased and Efficient Self-Supervised Incremental Contrastive Learning [31.763904668737304]
  We propose a self-supervised Incremental Contrastive Learning (ICL) framework consisting of a novel Incremental InfoNCE (NCE-II) loss function. ICL achieves up to 16.7x training speedup and 16.8x faster convergence with competitive results.
  arXiv  Detail & Related papers  (2023-01-28T06:11:31Z)
• False Correlation Reduction for Offline Reinforcement Learning [115.11954432080749]
  We propose falSe COrrelation REduction (SCORE) for offline RL, a practically effective and theoretically provable algorithm. We empirically show that SCORE achieves the SoTA performance with 3.1x acceleration on various tasks in a standard benchmark (D4RL)
  arXiv  Detail & Related papers  (2021-10-24T15:34:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.