Emotion Recognition in Conversation (ERC) is a crucial task for understanding human emotions and enabling natural human-computer interaction. Although Large Language Models (LLMs) have recently shown great potential in this field, their ability to capture the intrinsic connections between explicit and implicit emotions remains limited. We propose a novel ERC training framework, PRC-Emo, which integrates Prompt engineering, demonstration Retrieval, and Curriculum learning, with the goal of exploring whether LLMs can effectively perceive emotions in conversational contexts. Specifically, we design emotion-sensitive prompt templates based on both explicit and implicit emotional cues to better guide the model in understanding the speaker's psychological states. We construct the first dedicated demonstration retrieval repository for ERC, which includes training samples from widely used datasets, as well as high-quality dialogue examples generated by LLMs and manually verified. Moreover, we introduce a curriculum learning strategy into the LoRA fine-tuning process, incorporating weighted emotional shifts between same-speaker and different-speaker utterances to assign difficulty levels to dialogue samples, which are then organized in an easy-to-hard training sequence. Experimental results on two benchmark datasets -- IEMOCAP and MELD -- show that our method achieves new state-of-the-art (SOTA) performance, demonstrating the effectiveness and generalizability of our approach in improving LLM-based emotional understanding.

Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.

Abstract--Designing high-performance kernels requires expert-level tuning and a deep understanding of hardware characteristics. Recent advances in large language models (LLMs) have enabled automated kernel generation, yet most existing systems rely solely on correctness or execution time feedback, lacking the ability to reason about low-level performance bottlenecks. In this paper, we introduce PRAGMA, a profile-guided AI kernel generation framework that integrates execution feedback and fine-grained hardware profiling into the reasoning loop. PRAGMA enables LLMs to identify performance bottlenecks, preserve historical best versions, and iteratively refine code quality. Results show that PRAGMA consistently outperforms baseline N-PRAGMA without profiling enabled and achieves 2.81 and 2.30 averaged speedups against T orch on CPU and GPU platforms, respectively. Optimizing computational kernels is fundamental to achieving high performance in modern AI and HPC systems. Traditionally, reaching near-peak efficiency has required extensive manual tuning and deep expertise in architecture-specific optimization, making the development and maintenance of high-performance kernels both labor-intensive and error-prone.

Recovering meaningful concepts from language model activations is a central aim of interpretability. While existing feature extraction methods aim to identify concepts that are independent directions, it is unclear if this assumption can capture the rich temporal structure of language. Specifically, via a Bayesian lens, we demonstrate that Sparse Autoencoders (SAEs) impose priors that assume independence of concepts across time, implying stationarity. Meanwhile, language model representations exhibit rich temporal dynamics, including systematic growth in conceptual dimensionality, context-dependent correlations, and pronounced non-stationarity, in direct conflict with the priors of SAEs. Taking inspiration from computational neuroscience, we introduce a new interpretability objective -- Temporal Feature Analysis -- which possesses a temporal inductive bias to decompose representations at a given time into two parts: a predictable component, which can be inferred from the context, and a residual component, which captures novel information unexplained by the context. Temporal Feature Analyzers correctly parse garden path sentences, identify event boundaries, and more broadly delineate abstract, slow-moving information from novel, fast-moving information, while existing SAEs show significant pitfalls in all the above tasks. Overall, our results underscore the need for inductive biases that match the data in designing robust interpretability tools.

We investigate the performance of large language models on repetitive deterministic prediction tasks and study how the sequence accuracy rate scales with output length. Each such task involves repeating the same operation n times. Examples include letter replacement in strings following a given rule, integer addition, and multiplication of string operators in many body quantum mechanics. If the model performs the task through a simple repetition algorithm, the success rate should decay exponentially with sequence length. In contrast, our experiments on leading large language models reveal a sharp double exponential drop beyond a characteristic length scale, forming an accuracy cliff that marks the transition from reliable to unstable generation. This indicates that the models fail to execute each operation independently. To explain this phenomenon, we propose a statistical physics inspired model that captures the competition between external conditioning from the prompt and internal interference among generated tokens. The model quantitatively reproduces the observed crossover and provides an interpretable link between attention induced interference and sequence level failure. Fitting the model to empirical results across multiple models and tasks yields effective parameters that characterize the intrinsic error rate and error accumulation factor for each model task pair, offering a principled framework for understanding the limits of deterministic accuracy in large language models.

In agentic LLM scenarios, an agent's interaction process during a single rollout often exhibits branching behaviors. Due to memory retrieval and concurrent tool executions at certain decision points, the token trajectory of one task evolves into a tree-like structure rather than a linear sequence. However, current training pipelines decompose such tree-structured trajectories into separate linear segments, treating each branch as an independent sequence. As a result, shared prefixes across these branches are repeatedly recomputed during both forward and backward passes. To address this inefficiency, we propose Tree Training, a paradigm that computes each shared prefix only once and reuses its intermediate results across related branches during both forward and backward passes, substantially improving computation efficiency in large-scale agentic training. This is achieved via (i) Tree Packing, which efficiently reuses shared computations across trajectories, and (ii) Gradient Restoration, which ensures correct gradient propagation across reused prefixes. Experiments on multiple open-source models demonstrate up to 3.9x reduction in total training time, enabling more efficient agentic LLM SFT and RL training.

Millions of users now design personalized LLM-based chatbots that shape their daily interactions, yet they can only roughly anticipate how their design choices will manifest as behaviors in deployment. This opacity is consequential: seemingly innocuous prompts can trigger excessive sycophancy, toxicity, or other undesirable traits, degrading utility and raising safety concerns. To address this issue, we introduce an interface that enables neural transparency by exposing language model internals during chatbot design. Our approach extracts behavioral trait vectors (empathy, toxicity, sycophancy, etc.) by computing differences in neural activations between contrastive system prompts that elicit opposing behaviors. We predict chatbot behaviors by projecting the system prompt's final token activations onto these trait vectors, normalizing for cross-trait comparability, and visualizing results via an interactive sunburst diagram. To evaluate this approach, we conducted an online user study using Prolific to compare our neural transparency interface against a baseline chatbot interface without any form of transparency. Our analyses suggest that users systematically miscalibrated AI behavior: participants misjudged trait activations for eleven of fifteen analyzable traits, motivating the need for transparency tools in everyday human-AI interaction. While our interface did not change design iteration patterns, it significantly increased user trust and was enthusiastically received. Qualitative analysis revealed nuanced user experiences with the visualization, suggesting interface and interaction improvements for future work. This work offers a path for how mechanistic interpretability can be operationalized for non-technical users, establishing a foundation for safer, more aligned human-AI interactions.

Coral reefs are rapidly declining due to anthropogenic pressures such as climate change, underscoring the urgent need for scalable, automated monitoring. We introduce ReefNet, a large public coral reef image dataset with point-label annotations mapped to the World Register of Marine Species (WoRMS). ReefNet aggregates imagery from 76 curated CoralNet sources and an additional site from Al Wajh in the Red Sea, totaling approximately 925000 genus-level hard coral annotations with expert-verified labels. Unlike prior datasets, which are often limited by size, geography, or coarse labels and are not ML-ready, ReefNet offers fine-grained, taxonomically mapped labels at a global scale to WoRMS. We propose two evaluation settings: (i) a within-source benchmark that partitions each source's images for localized evaluation, and (ii) a cross-source benchmark that withholds entire sources to test domain generalization. We analyze both supervised and zero-shot classification performance on ReefNet and find that while supervised within-source performance is promising, supervised performance drops sharply across domains, and performance is low across the board for zero-shot models, especially for rare and visually similar genera. This provides a challenging benchmark intended to catalyze advances in domain generalization and fine-grained coral classification. We will release our dataset, benchmarking code, and pretrained models to advance robust, domain-adaptive, global coral reef monitoring and conservation.

The core premise of AI debate as a scalable oversight technique is that it is harder to lie convincingly than to refute a lie, enabling the judge to identify the correct position. Yet, existing debate experiments have relied on datasets with ground truth, where lying is reduced to defending an incorrect proposition. This overlooks a subjective dimension: lying also requires the belief that the claim defended is false. In this work, we apply debate to subjective questions and explicitly measure large language models' prior beliefs before experiments. Debaters were asked to select their preferred position, then presented with a judge persona deliberately designed to conflict with their identified priors. This setup tested whether models would adopt sycophantic strategies, aligning with the judge's presumed perspective to maximize persuasiveness, or remain faithful to their prior beliefs. We implemented and compared two debate protocols, sequential and simultaneous, to evaluate potential systematic biases. Finally, we assessed whether models were more persuasive and produced higher-quality arguments when defending positions consistent with their prior beliefs versus when arguing against them. Our main findings show that models tend to prefer defending stances aligned with the judge persona rather than their prior beliefs, sequential debate introduces significant bias favoring the second debater, models are more persuasive when defending positions aligned with their prior beliefs, and paradoxically, arguments misaligned with prior beliefs are rated as higher quality in pairwise comparison. These results can inform human judges to provide higher-quality training signals and contribute to more aligned AI systems, while revealing important aspects of human-AI interaction regarding persuasion dynamics in language models.

Psychological assessments are dominated by rating scales, which cannot capture the nuance in natural language. Efforts to supplement them with qualitative text have relied on labelled datasets or expert rubrics, limiting scalability. We introduce a framewo rk that avoids this reliance: large language models (LLMs) score free - text responses with simple prompts to produce candidate LLM items, from which we retain those that yield the most test information when co - calibrated with a baseline scale. Using depress ion as a case study, we developed and tested the method in upper - secondary students (n=693) and a matched synthetic dataset (n=3,000). Results on held - out test sets show ed that augmenting a 19 - item scale with LLM items improved its precision, accuracy, and convergent validity. Further, the test information gain matched that of adding as many as 16 rating - scale items. This framework leverage s the increas ing availability of transcribed language to enhance psychometric measures, with applications in clinical h ealth and beyond.