We introduce CLEAR-3K, a dataset of 3,000 assertion-reasoning questions designed to evaluate whether language models can determine if one statement causally explains another. Each question present an assertion-reason pair and challenge language models to distinguish between semantic relatedness and genuine causal explanatory relationships. Through comprehensive evaluation of 21 state-of-the-art language models (ranging from 0.5B to 72B parameters), we identify two fundamental findings. First, language models frequently confuse semantic similarity with causality, relying on lexical and semantic overlap instead of inferring actual causal explanatory relationships. Second, as parameter size increases, models tend to shift from being overly skeptical about causal relationships to being excessively permissive in accepting them. Despite this shift, performance measured by the Matthews Correlation Coefficient plateaus at just 0.55, even for the best-performing models.Hence, CLEAR-3K provides a crucial benchmark for developing and evaluating genuine causal reasoning in language models, which is an essential capability for applications that require accurate assessment of causal relationships.

Clinical robustness is critical to the safe deployment of medical Large Language Models (LLMs), but key questions remain about how LLMs and humans may differ in response to the real-world variability typified by clinical settings. To address this, we introduce MedPerturb, a dataset designed to systematically evaluate medical LLMs under controlled perturbations of clinical input. MedPerturb consists of clinical vignettes spanning a range of pathologies, each transformed along three axes: (1) gender modifications (e.g., gender-swapping or gender-removal); (2) style variation (e.g., uncertain phrasing or colloquial tone); and (3) format changes (e.g., LLM-generated multi-turn conversations or summaries). With MedPerturb, we release a dataset of 800 clinical contexts grounded in realistic input variability, outputs from four LLMs, and three human expert reads per clinical context. We use MedPerturb in two case studies to reveal how shifts in gender identity cues, language style, or format reflect diverging treatment selections between humans and LLMs. We find that LLMs are more sensitive to gender and style perturbations while human annotators are more sensitive to LLM-generated format perturbations such as clinical summaries. Our results highlight the need for evaluation frameworks that go beyond static benchmarks to assess the similarity between human clinician and LLM decisions under the variability characteristic of clinical settings.

Large language models (LLMs) are increasingly used to generate feedback, yet their impact on learning remains underexplored, especially compared to existing feedback methods. This study investigates how on-demand LLM-generated explanatory feedback influences learning in seven scenario-based tutor training lessons. Analyzing over 2,600 lesson completions from 885 tutor learners, we compare posttest performance among learners across three groups: learners who received feedback generated by gpt-3.5-turbo, those who declined it, and those without access. All groups received non-LLM corrective feedback. To address potential selection bias-where higher-performing learners may be more inclined to use LLM feedback-we applied propensity scoring. Learners with a higher predicted likelihood of engaging with LLM feedback scored significantly higher at posttest than those with lower propensity. After adjusting for this effect, two out of seven lessons showed statistically significant learning benefits from LLM feedback with standardized effect sizes of 0.28 and 0.33. These moderate effects suggest that the effectiveness of LLM feedback depends on the learners' tendency to seek support. Importantly, LLM feedback did not significantly increase completion time, and learners overwhelmingly rated it as helpful. These findings highlight LLM feedback's potential as a low-cost and scalable way to improve learning on open-ended tasks, particularly in existing systems already providing feedback without LLMs. This work contributes open datasets, LLM prompts, and rubrics to support reproducibility.

Accurately assessing student knowledge is critical for effective education, yet traditional Knowledge Tracing (KT) methods rely on opaque latent embeddings, limiting interpretability. Even LLM-based approaches generate direct predictions or summaries that may hallucinate without any accuracy guarantees. We recast KT as an inverse problem: learning the minimum natural-language summary that makes past answers explainable and future answers predictable. Our Language Bottleneck Model (LBM) consists of an encoder LLM that writes an interpretable knowledge summary and a frozen decoder LLM that must reconstruct and predict student responses using only that summary text. By constraining all predictive information to pass through a short natural-language bottleneck, LBMs ensure that the summary contains accurate information while remaining human-interpretable. Experiments on synthetic arithmetic benchmarks and the large-scale Eedi dataset show that LBMs rival the accuracy of state-of-the-art KT and direct LLM methods while requiring orders-of-magnitude fewer student trajectories. We demonstrate that training the encoder with group-relative policy optimization, using downstream decoding accuracy as a reward signal, effectively improves summary quality.

With the rapid progress of speech language models (SLMs), discrete speech tokens have emerged as a core interface between speech and text, enabling unified modeling across modalities. Recent speech tokenization approaches aim to isolate semantic information from low-level acoustics to better align with language models. In particular, previous methods use SSL teachers such as HuBERT to extract semantic representations, which are then distilled into a semantic quantizer to suppress acoustic redundancy as well as capture content-related latent structures. However, they still produce speech token sequences significantly longer than their textual counterparts, creating challenges for efficient speech-language modeling. Reducing the frame rate is a natural solution, but standard techniques, such as rigid average pooling across frames, can distort or dilute the semantic structure required for effective LM alignment. To address this, we propose LM-SPT, a speech tokenization method that introduces a novel semantic distillation. Instead of directly matching teacher and student features via pooling, we reconstruct speech solely from semantic tokens and minimize the discrepancy between the encoded representations of the original and reconstructed waveforms, obtained from a frozen automatic speech recognition (ASR) encoder. This indirect yet data-driven supervision enables the tokenizer to learn discrete units that are more semantically aligned with language models. LM-SPT further incorporates architectural improvements to the encoder and decoder for speech tokenization, and supports multiple frame rates, including 25Hz, 12.5Hz, and 6.25Hz. Experimental results show that LM-SPT achieves superior reconstruction fidelity compared to baselines, and that SLMs trained with LM-SPT tokens achieve competitive performances on speech-to-text and consistently outperform baselines on text-to-speech tasks.

Large language models (LLMs) are rapidly being adopted across psychology, serving as research tools, experimental subjects, human simulators, and computational models of cognition. However, the application of human measurement tools to these systems can produce contradictory results, raising concerns that many findings are measurement phantoms--statistical artifacts rather than genuine psychological phenomena. In this Perspective, we argue that building a robust science of AI psychology requires integrating two of our field's foundational pillars: the principles of reliable measurement and the standards for sound causal inference. We present a dual-validity framework to guide this integration, which clarifies how the evidence needed to support a claim scales with its scientific ambition. Using an LLM to classify text may require only basic accuracy checks, whereas claiming it can simulate anxiety demands a far more rigorous validation process. Current practice systematically fails to meet these requirements, often treating statistical pattern matching as evidence of psychological phenomena. The same model output--endorsing "I am anxious"--requires different validation strategies depending on whether researchers claim to measure, characterize, simulate, or model psychological constructs. Moving forward requires developing computational analogues of psychological constructs and establishing clear, scalable standards of evidence rather than the uncritical application of human measurement tools.

Reinforcement learning algorithms often suffer from poor sample efficiency, making them challenging to apply in multi-task or continual learning settings. Efficiency can be improved by transferring knowledge from a previously trained teacher policy to guide exploration in new but related tasks. However, if the new task sufficiently differs from the teacher's training task, the transferred guidance may be sub-optimal and bias exploration toward low-reward behaviors. We propose an energy-based transfer learning method that uses out-of-distribution detection to selectively issue guidance, enabling the teacher to intervene only in states within its training distribution. We theoretically show that energy scores reflect the teacher's state-visitation density and empirically demonstrate improved sample efficiency and performance across both single-task and multi-task settings.

Understanding whether large language models (LLMs) possess a world model-a structured understanding of the world that supports generalization beyond surface-level patterns-is central to assessing their reliability, especially in high-stakes applications. We propose a formal framework for evaluating whether an LLM exhibits a sufficiently robust world model, defined as producing consistent outputs across semantically equivalent prompts while distinguishing between prompts that express different intents. We introduce a new evaluation approach to measure this that decomposes model response variability into three components: variability due to user purpose, user articulation, and model instability. An LLM with a strong world model should attribute most of the variability in its responses to changes in foundational purpose rather than superficial changes in articulation. This approach allows us to quantify how much of a model's behavior is semantically grounded rather than driven by model instability or alternative wording. We apply this framework to evaluate LLMs across diverse domains. Our results show how larger models attribute a greater share of output variability to changes in user purpose, indicating a more robust world model. This improvement is not uniform, however: larger models do not consistently outperform smaller ones across all domains, and their advantage in robustness is often modest. These findings highlight the importance of moving beyond accuracy-based benchmarks toward semantic diagnostics that more directly assess the structure and stability of a model's internal understanding of the world.

Reinforcement learning (RL) exhibits remarkable potential in addressing autonomous driving tasks. However, it is difficult to train a sample-efficient and safe policy in complex scenarios. In this article, we propose a novel hierarchical reinforcement learning (HRL) framework with a goal-conditioned collision prediction (GCCP) module. In the hierarchical structure, the GCCP module predicts collision risks according to different potential subgoals of the ego vehicle. A high-level decision-maker choose the best safe subgoal. A low-level motion-planner interacts with the environment according to the subgoal. Compared to traditional RL methods, our algorithm is more sample-efficient, since its hierarchical structure allows reusing the policies of subgoals across similar tasks for various navigation scenarios. In additional, the GCCP module's ability to predict both the ego vehicle's and surrounding vehicles' future actions according to different subgoals, ensures the safety of the ego vehicle throughout the decision-making process. Experimental results demonstrate that the proposed method converges to an optimal policy faster and achieves higher safety than traditional RL methods.

Data plays the most prominent role in how language models acquire skills and knowledge. The lack of massive, well-organized pre-training datasets results in costly and inaccessible data pipelines. We present Essential-Web v1.0, a 24-trillion-token dataset in which every document is annotated with a twelve-category taxonomy covering topic, format, content complexity, and quality. Taxonomy labels are produced by EAI-Distill-0.5b, a fine-tuned 0.5b-parameter model that achieves an annotator agreement within 3% of Qwen2.5-32B-Instruct. With nothing more than SQL-style filters, we obtain competitive web-curated datasets in math (-8.0% relative to SOTA), web code (+14.3%), STEM (+24.5%) and medical (+8.6%). Essential-Web v1.0 is available on HuggingFace: https://huggingface.co/datasets/EssentialAI/essential-web-v1.0