Recent work reports that Large Reasoning Models (LRMs) undergo a collapse in performance on solving puzzles beyond certain perplexity thresholds. In subsequent discourse, questions have arisen as to whether the nature of the task muddles an evaluation of true reasoning. One potential confound is the requirement that the model keep track of the state space on its own. We provide a large language model (LLM) with an environment interface for Tower of Hanoi problems, allowing it to make a move with a tool call, provide written justification, observe the resulting state space, and reprompt itself for the next move. We observe that access to an environment interface does not delay or eradicate performance collapse. Furthermore, LLM-parameterized policy analysis reveals increasing divergence from both optimal policies and uniformly random policies, suggesting that the model exhibits mode-like collapse at each level of complexity, and that performance is dependent upon whether the mode reflects the correct solution for the problem. We suggest that a similar phenomena might take place in LRMs.

Recent successes of reinforcement learning (RL) in training large reasoning models motivate the question of whether self-training - the process where a model learns from its own judgments - can be sustained within RL. In this work, we study this question using majority voting as a simple self-feedback mechanism. On a comprehensive set of experiments on both synthetic and real reasoning tasks, we find that this basic approach improves not only the model's reasoning performance, but also its capability of generating better quality feedback for the next RL iteration, driving further model improvement. Yet our analysis also reveals a critical limitation of such a self-training paradigm - prolonged RL with self-reward leads to reward hacking where models learn to maximize training (pseudo-)reward, resulting in sudden and complete performance collapse. Together, these results highlight feedback design as the central challenge and call for future research on mechanisms to enable prolonged self-improvement.

We show that the proposed method significantly improves the performance of the baseline algorithm in the considered environments.

The recent work by Shojaee et al. (2025), titled The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity, presents a compelling empirical finding, a reasoning cliff, where the performance of Large Reasoning Models (LRMs) collapses beyond a specific complexity threshold, which the authors posit as an intrinsic scaling limitation of Chain-of-Thought (CoT) reasoning. This commentary, while acknowledging the study's methodological rigor, contends that this conclusion is confounded by experimental artifacts. We argue that the observed failure is not evidence of a fundamental cognitive boundary, but rather a predictable outcome of system-level constraints in the static, text-only evaluation paradigm, including tool use restrictions, context window recall issues, the absence of crucial cognitive baselines, inadequate statistical reporting, and output generation limits. We reframe this performance collapse through the lens of an agentic gap, asserting that the models are not failing at reasoning, but at execution within a profoundly restrictive interface. We empirically substantiate this critique by demonstrating a striking reversal. A model, initially declaring a puzzle impossible when confined to text-only generation, now employs agentic tools to not only solve it but also master variations of complexity far beyond the reasoning cliff it previously failed to surmount. Additionally, our empirical analysis of tool-enabled models like o4-mini and GPT-4o reveals a hierarchy of agentic reasoning, from simple procedural execution to complex meta-cognitive self-correction, which has significant implications for how we define and measure machine intelligence. The illusion of thinking attributed to LRMs is less a reasoning deficit and more a consequence of an otherwise capable mind lacking the tools for action.

Neural Architecture Search (NAS) currently relies heavily on labeled data, which is both expensive and time-consuming to acquire. In this paper, we propose a novel NAS framework based on Masked Autoencoders (MAE) that eliminates the need for labeled data during the search process. By replacing the supervised learning objective with an image reconstruction task, our approach enables the robust discovery of network architectures without compromising performance and generalization ability. Additionally, we address the problem of performance collapse encountered in the widely-used Differentiable Architecture Search (DARTS) method in the unsupervised paradigm by introducing a multi-scale decoder. Through extensive experiments conducted on various search spaces and datasets, we demonstrate the effectiveness and robustness of the proposed method, providing empirical evidence of its superiority over baseline approaches.

Deep reinforcement learning (RL) has achieved remarkable success in solving complex tasks through its integration with deep neural networks (DNNs) as function approximators. However, the reliance on DNNs has introduced a new challenge called primacy bias, whereby these function approximators tend to prioritize early experiences, leading to overfitting. To mitigate this primacy bias, a reset method has been proposed, which performs periodic resets of a portion or the entirety of a deep RL agent while preserving the replay buffer. However, the use of the reset method can result in performance collapses after executing the reset, which can be detrimental from the perspective of safe RL and regret minimization. In this paper, we propose a new reset-based method that leverages deep ensemble learning to address the limitations of the vanilla reset method and enhance sample efficiency. The proposed method is evaluated through various experiments including those in the domain of safe RL. Numerical results show its effectiveness in high sample efficiency and safety considerations.

Differentiable neural architecture search (DARTS) is a popular method for neural architecture search (NAS), which performs cell-search and utilizes continuous relaxation to improve the search efficiency via gradient-based optimization. The main shortcoming of DARTS is performance collapse, where the discovered architecture suffers from a pattern of declining quality during search. Performance collapse has become an important topic of research, with many methods trying to solve the issue through either regularization or fundamental changes to DARTS. However, the weight-sharing framework used for cell-search in DARTS and the convergence of architecture parameters has not been analyzed yet. In this paper, we provide a thorough and novel theoretical and empirical analysis on DARTS and its point of convergence. We show that DARTS suffers from a specific structural flaw due to its weight-sharing framework that limits the convergence of DARTS to saturation points of the softmax function. This point of convergence gives an unfair advantage to layers closer to the output in choosing the optimal architecture, causing performance collapse. We then propose two new regularization terms that aim to prevent performance collapse by harmonizing operation selection via aligning gradients of layers. Experimental results on six different search spaces and three different datasets show that our method ($\Lambda$-DARTS) does indeed prevent performance collapse, providing justification for our theoretical analysis and the proposed remedy.

Offline reinforcement learning, by learning from a fixed dataset, makes it possible to learn agent behaviors without interacting with the environment. However, depending on the quality of the offline dataset, such pre-trained agents may have limited performance and would further need to be fine-tuned online by interacting with the environment. During online fine-tuning, the performance of the pre-trained agent may collapse quickly due to the sudden distribution shift from offline to online data. While constraints enforced by offline RL methods such as a behaviour cloning loss prevent this to an extent, these constraints also significantly slow down online fine-tuning by forcing the agent to stay close to the behavior policy. We propose to adaptively weigh the behavior cloning loss during online fine-tuning based on the agent's performance and training stability. Moreover, we use a randomized ensemble of Q functions to further increase the sample efficiency of online fine-tuning by performing a large number of learning updates. Experiments show that the proposed method yields state-of-the-art offline-to-online reinforcement learning performance on the popular D4RL benchmark. Code is available: \url{https://github.com/zhaoyi11/adaptive_bc}.