Reasoning-focused LLMs sometimes alter their behavior when they detect that they are being evaluated--which can lead them to optimize for test-passing performance or to comply more readily with harmful prompts if real-world consequences appear absent. We present the first quantitative study of how such "test awareness" impacts model behavior, particularly its performance on safety-related tasks1. We introduce a white-box probing framework that (i) linearly identifies awareness-related activations and (ii) steers models toward or away from test awareness while monitoring downstream performance. We apply our method to different state-of-the-art openweight reasoning LLMs across both realistic and hypothetical tasks (denoting tests or simulations). Our results demonstrate that test awareness significantly impacts safety alignment (such as compliance with harmful requests and conforming to stereotypes) with effects varying in both magnitude and direction across models. By providing control over this latent effect, our work aims to provide a stress-test mechanism and increase trust in how we perform safety evaluations.

When a vision model performs image recognition, which visual attributes drive its predictions? Detecting unintended reliance on specific visual features is critical for ensuring model robustness, preventing overfitting, and avoiding spurious correlations. We introduce an automated framework for detecting such dependencies in trained vision models. At the core of our method is a self-reflective agent that systematically generates and tests hypotheses about visual attributes that a model may rely on. This process is iterative: the agent refines its hypotheses based on experimental outcomes and uses a self-evaluation protocol to assess whether its findings accurately explain model behavior. When inconsistencies arise, the agent self-reflects over its findings and triggers a new cycle of experimentation. We evaluate our approach on a novel benchmark of 130 models designed to exhibit diverse visual attribute dependencies across 18 categories. Our results show that the agent's performance consistently improves with self-reflection, with a significant performance increase over non-reflective baselines. We further demonstrate that the agent identifies real-world visual attribute dependencies in state-of-the-art models, including CLIP's vision encoder and the YOLOv8 object detector.

Transformer models have driven breakthroughs across various language tasks by their strong capability to learn rich contextual representations. Scaling them to improve representation, however, often demands substantial memory and compute costs, such as the Key-Value (KV) cache used during auto-regressive decoding. Skip connections offer a promising way to improve representation without bloating resource usage, yet most prior works either improve expressivity while leaving KV costs unchanged, or reduce memory at the cost of weaker representation. In this work, we propose SkipV1Former, a Transformer variant that uses skip connections from the first layer's Value heads to strengthen model representation and reduce KV cache. Specifically, from the second block onward, each layer reuses half of its Value heads from the very first layer, while computing the other half as usualcutting Value projections and V cache by nearly 50 %.

Pre-trained diffusion models are commonly used to generate clean data (e.g., images) from random noises, effectively forming pairs of noises and corresponding clean images. Distillation on these pre-trained models can be viewed as the process of constructing advanced trajectories within the pair to accelerate sampling. For instance, consistency model distillation develops consistent projection functions to regulate trajectories, although sampling efficiency remains a concern. Rectified flow method enforces straight trajectories to enable faster sampling, yet relies on numerical ODE solvers, which may introduce approximation errors. In this work, we bridge the gap between the consistency model and the rectified flow method by proposing a Straight-Consistent Trajectories (SCoT) model. SCoT enjoys the benefits of both approaches for fast sampling, producing trajectories with consistent and straight properties simultaneously. These dual properties are strategically balanced by targeting two critical objectives: (1) regulating the gradient of SCoT's mapping function to a constant and (2) ensuring trajectory consistency. Extensive experimental results demonstrate the effectiveness and efficiency of SCoT.

Parameter-efficient fine-tuning (PEFT) is essential for adapting large foundation models without excessive storage cost. However, current approaches such as LoRA treat each layer's adaptation independently, overlooking correlations across layers. This independence causes the number of trainable parameters to grow linearly with model depth. We provide theoretical and empirical evidence that skip connections in transformers create smooth gradient propagation across layers. This smoothness leads to weight adaptations that concentrate most of their energy in low-frequency spectral components, especially along the layer dimension.

Diffusion models have achieved remarkable success across a wide range of generative tasks. A key challenge is understanding the mechanisms that prevent their memorization of training data and allow generalization. In this work, we investigate the role of the training dynamics in the transition from generalization to memorization. Through extensive experiments and theoretical analysis, we identify two distinct timescales: an early time τgen at which models begin to generate high-quality samples, and a later time τmem beyond which memorization emerges. Crucially, we find that τmem increases linearly with the training set size n, while τgen remains constant. This creates a growing window of training times with n where models generalize effectively, despite showing strong memorization if training continues beyond it. It is only when nbecomes larger than a model-dependent threshold that overfitting disappears at infinite training times. These findings reveal a form of implicit dynamical regularization in the training dynamics, which allow to avoid memorization even in highly overparameterized settings. Our results are supported by numerical experiments with standard U-Net architectures on realistic and synthetic datasets, and by a theoretical analysis using a tractable random features model studied in the high-dimensional limit.

The rapid emergence of diverse large language models (LLMs) has spurred the development of LLM routers that assign user queries to the most suitable model. However, existing LLM routers typically perform a single-round, one-to-one mapping (i.e., assigning each query to a single model in isolation), which limits their capability to tackle complex tasks that demand the complementary strengths of multiple LLMs. In this paper, we present Router-R1, a reinforcement learning (RL)-based framework that formulates multi-LLM routing and aggregation as a sequential decision process. Router-R1 instantiates the router itself as a capable LLM, leveraging its reasoning ability to interleave "think" actions (internal deliberation) with "route" actions (dynamic model invocation), and integrates each response into its evolving context. To facilitate learning, we employ a lightweight rule-based reward comprising format rewards, final outcome rewards, and a novel cost reward for optimizing the balance between performance and cost, opening a pathway toward enhancing performance-cost trade-offs via RL.

The core task of recommender systems is to learn user preferences from historical user-item interactions. With the rapid development of large language models (LLMs), recent research has explored leveraging the reasoning capabilities of LLMs to enhance rating prediction tasks. However, existing distillation-based methods suffer from limitations such as the teacher model's insufficient recommendation capability, costly and static supervision, and superficial transfer of reasoning ability. To address these issues, this paper proposes RecZero, a reinforcement learning (RL)-based recommendation paradigm that abandons the traditional multi-model and multi-stage distillation approach. Instead, RecZero trains a single LLM through pure RL to autonomously develop reasoning capabilities for rating prediction. RecZero consists of two key components: (1) "Think-before-Recommendation" prompt construction, which employs a structured reasoning template to guide the model in step-wise analysis of user interests, item features, and user-item compatibility; and (2) rule-based reward modeling, which adopts group relative policy optimization (GRPO) to compute rewards for reasoning trajectories and optimize the LLM. Additionally, the paper explores a hybrid paradigm, RecOne, which combines supervised fine-tuning with RL, initializing the model with coldstart reasoning samples and further optimizing it with RL. Experimental results demonstrate that RecZero and RecOne significantly outperform existing baseline methods on multiple benchmark datasets, validating the superiority of the RL paradigm in achieving autonomous reasoning-enhanced recommender systems.

High-quality supervised finetuning (SFT) data are essential for unlocking pretrained LLMs' capabilities. Typically, instructions are paired with responses from various sources by humans annotators or other LMs, which are often out of the distribution of the target model to be finetuned. This, at scale, can lead to diminishing returns and even hurt the models' performance and robustness. We hypothesize that SFT is most effective with data aligned to the model's pretrained distribution and propose GRAPE-- a novel SFT framework that tailors supervision to the target model. For each instruction, it gathers responses from various sources, and selects the one that aligns most closely to the target model's pretrained distribution, as measured by the normalized probability. We then proceed with standard SFT with these selected responses. We first evaluate GRAPE with a controlled experiment, where we sample various solutions for each question in UltraInteract from multiple models and finetune on GRAPE-selected data using LMs from different families including LLaMA.1-8B,

Misinformation detectors often rely on superficial cues (i.e., shortcuts) that correlate with misinformation in training data but fail to generalize to the diverse and evolving nature of real-world misinformation. This issue is exacerbated by large language models (LLMs), which can easily generate convincing misinformation using simple prompts. We introduce TRUTHOVERTRICKS, a unified evaluation paradigm for measuring shortcut learning in misinformation detection. TRUTHOVERTRICKS categorizes shortcut behaviors into intrinsic shortcut induction and extrinsic shortcut injection, and evaluates seven representative detectors across 14 popular benchmarks, along with two new factual misinformation datasets, NQ-Misinfo and Streaming-Misinfo. Empirical results reveal that existing detectors suffer severe performance degradation when exposed to both naturally occurring and adversarially crafted shortcuts. To address this, we propose the Shortcut Mitigation Framework (SMF), an LLM-augmented data augmentation framework that mitigates shortcut reliance through paraphrasing, factual summarization, and sentiment normalization. SMF consistently enhances robustness across 16 benchmarks, forcing models to rely on deeper semantic understanding rather than shortcut cues.