mlsanityData to 14 Oct 2026

Do Not Imitate, Reinforce: Iterative Classification via Belief Refinement

arXiv 2604.22110 ↗First posted 23 Apr 2026cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Standard supervised classification trains models to imitate the exact labels provided by a perfect oracle. This imitation happens in a single pass, restricting the model to a fixed compute budget even when inputs vary in complexity. Moreover, the rigid training objective forces the model to express absolute certainty on its training data, resulting in overconfident predictions during evaluation. We propose Reinforced Iterative Classification (RIC), which replaces the imitative objective with Reinforcement Learning (RL). RIC deploys a recurrent agent that iteratively updates a predictive distribution over classes, receiving reward for stepwise improvement in prediction quality. The value function provides a natural halting criterion by estimating the remaining scope for improvement. We prove that the iterative formulation recovers the same optimal predictions as cross-entropy while yielding an anytime classifier. On image classification benchmarks, RIC matches the accuracy of supervised baselines with improved calibration and learns to allocate computation adaptively across inputs.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 26 of its 32 references that are in the map. First posted on arXiv 23 Apr 2026.

Builds on

26 of 32 references.

References · 26

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  2. 2017On Calibration of Modern Neural Networkscited by 4,960
  3. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  4. 2018Evidential Deep Learning to Quantify Classification Uncertaintycited by 732
  5. 2017The Implicit Bias of Gradient Descent on Separable Datacited by 655
  6. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  7. 2014Deeply-Supervised Netscited by 367
  8. 2014Recurrent Models of Visual Attentioncited by 349
  9. 2017SkipNet: Learning Dynamic Routing in Convolutional Networkscited by 332
  10. 2021Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotationscited by 256
  11. 2022Mitigating Neural Network Overconfidence with Logit Normalizationcited by 252
  12. 2017BlockDrop: Dynamic Inference Paths in Residual Networkscited by 189
14 more
  1. 2021PonderNet: Learning to Pondercited by 168
  2. 2016Spatially Adaptive Computation Time for Residual Networkscited by 118
  3. 2025Beyond Binary Rewards: Training LMs to Reason About Their Uncertaintycited by 117
  4. 2024Taming Overconfidence in LLMs: Reward Calibration in RLHFcited by 112
  5. 2025Rewarding Doubt: A Reinforcement Learning Approach to Calibrated Confidence Expression of Large Language Modelscited by 52
  6. 2024Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?cited by 35
  7. 2024Simple Policy Optimizationcited by 32
  8. 2025Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoningcited by 29
  9. 2022Recurrent networks, hidden states and beliefs in partially observable environmentscited by 12
  10. 2019Multi-Agent Image Classification via Reinforcement Learningcited by 8
  11. 2025Learning When to Stop: Adaptive Latent Reasoning via Reinforcement Learningcited by 7
  12. 2025When Can Model-Free Reinforcement Learning be Enough for Thinking?cited by 6
  13. 2025Toward Agents That Reason About Their Computationcited by 3
  14. 2026Primate-like perceptual decision making emerges through deep recurrent reinforcement learningcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Prototype Latent World Model Replay for Class-Incremental Learning
  2. 2026LookThere! Sparse Vision by Reinforced Selection
  3. 2025Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL
  4. 2024Reinforcement Learning with Generative Models for Compact Support Sets
  5. 2026DocHRL: A Hierarchical Reinforcement Learning Framework for Cost-Optimised Document Classification
  6. 2026ClaHF: A Human Feedback-inspired Reinforcement Learning Framework for Improving Classification Tasks
  7. 2020Active Imitation Learning with Noisy Guidance
  8. 2025IMPROVE: Iterative Model Pipeline Refinement and Optimization Leveraging LLM Experts