Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.

As Large Language Models (LLMs) are increasingly deployed in sensitive domains, traditional data privacy measures prove inadequate for protecting information that is implicit, contextual, or inferable - what we define as semantic privacy. This Systematization of Knowledge (SoK) introduces a lifecycle-centric framework to analyze how semantic privacy risks emerge across input processing, pretraining, fine-tuning, and alignment stages of LLMs. We categorize key attack vectors and assess how current defenses, such as differential privacy, embedding encryption, edge computing, and unlearning, address these threats. Our analysis reveals critical gaps in semantic-level protection, especially against contextual inference and latent representation leakage. We conclude by outlining open challenges, including quantifying semantic leakage, protecting multimodal inputs, balancing de-identification with generation quality, and ensuring transparency in privacy enforcement. This work aims to inform future research on designing robust, semantically aware privacy-preserving techniques for LLMs.

With the rapid development of digital technology, AI-driven psychological counseling has gradually become an important research direction in the field of mental health. However, existing models still have deficiencies in dialogue safety, detailed scenario handling, and lightweight deployment. To address these issues, this study proposes PsyLite, a lightweight psychological counseling large language model agent developed based on the base model InternLM2.5-7B-chat. Through a two-stage training strategy (hybrid distillation data fine-tuning and ORPO preference optimization), PsyLite enhances the model's deep-reasoning ability, psychological counseling ability, and safe dialogue ability. After deployment using Ollama and Open WebUI, a custom workflow is created with Pipelines. An innovative conditional RAG is designed to introduce crosstalk humor elements at appropriate times during psychological counseling to enhance user experience and decline dangerous requests to strengthen dialogue safety. Evaluations show that PsyLite outperforms the baseline models in the Chinese general evaluation (CEval), psychological counseling professional evaluation (CPsyCounE), and dialogue safety evaluation (SafeDialBench), particularly in psychological counseling professionalism (CPsyCounE score improvement of 47.6\%) and dialogue safety (\safe{} score improvement of 2.4\%). Additionally, the model uses quantization technology (GGUF q4\_k\_m) to achieve low hardware deployment (5GB memory is sufficient for operation), providing a feasible solution for psychological counseling applications in resource-constrained environments.

The ability to perform reliable long-horizon task planning is crucial for deploying robots in real-world environments. However, directly employing Large Language Models (LLMs) as action sequence generators often results in low success rates due to their limited reasoning ability for long-horizon embodied tasks. In the STEP framework, we construct a subgoal tree through a pair of closed-loop models: a subgoal decomposition model and a leaf node termination model. Within this framework, we develop a hierarchical tree structure that spans from coarse to fine resolutions. The subgoal decomposition model leverages a foundation LLM to break down complex goals into manageable subgoals, thereby spanning the subgoal tree. The leaf node termination model provides real-time feedback based on environmental states, determining when to terminate the tree spanning and ensuring each leaf node can be directly converted into a primitive action. Experiments conducted in both the VirtualHome WAH-NL benchmark and on real robots demonstrate that STEP achieves long-horizon embodied task completion with success rates up to 34% (WAH-NL) and 25% (real robot) outperforming SOTA methods.

Training Large Language Models (LLMs) is prohibitively expensive, creating a critical scaling gap where insights from small-scale experiments often fail to transfer to resource-intensive production systems, thereby hindering efficient innovation. To bridge this, we introduce Farseer, a novel and refined scaling law offering enhanced predictive accuracy across scales. By systematically constructing a model loss surface $L(N,D)$, Farseer achieves a significantly better fit to empirical data than prior laws (e.g., Chinchilla's law). Our methodology yields accurate, robust, and highly generalizable predictions, demonstrating excellent extrapolation capabilities, improving upon Chinchilla's law by reducing extrapolation error by 433\%. This allows for the reliable evaluation of competing training strategies across all $(N,D)$ settings, enabling conclusions from small-scale ablation studies to be confidently extrapolated to predict large-scale performance. Furthermore, Farseer provides new insights into optimal compute allocation, better reflecting the nuanced demands of modern LLM training. To validate our approach, we trained an extensive suite of approximately 1,000 LLMs across diverse scales and configurations, consuming roughly 3 million NVIDIA H100 GPU hours. We are comprehensively open-sourcing all models, data, results, and logs at https://github.com/Farseer-Scaling-Law/Farseer to foster further research.

If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised. For example, models could have systematically different behavior during evaluations, leading to less reliable benchmarks for deployment and governance decisions. We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness. To achieve this, we construct a diverse benchmark of 1,000 prompts and transcripts from 61 distinct datasets. These span public benchmarks (e.g., MMLU, SWEBench), real-world deployment interactions, and agent trajectories from scaffolding frameworks (e.g., web-browsing agents). Frontier models clearly demonstrate above-random evaluation awareness (Gemini-2.5-Pro reaches an AUC of $0.83$), but do not yet surpass our simple human baseline (AUC of $0.92$). Furthermore, both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings. Additionally, we test whether models can identify the purpose of the evaluation. Under multiple-choice and open-ended questioning, AI models far outperform random chance in identifying what an evaluation is testing for. Our results indicate that frontier models already exhibit a substantial, though not yet superhuman, level of evaluation-awareness. We recommend tracking this capability in future models.

--Recent advancements in Large Language Models (LLMs) have transformed code generation from natural language queries. However, despite their extensive knowledge and ability to produce high-quality code, LLMs often struggle with contextual accuracy, particularly in evolving codebases. Current code search and retrieval methods frequently lack robustness in both the quality and contextual relevance of retrieved results, leading to suboptimal code generation. This paper introduces a novel knowledge graph-based approach to improve code search and retrieval leading to better quality of code generation in the context of repository-level tasks. The proposed approach represents code repositories as graphs, capturing structural and relational information for enhanced context-aware code generation. Our framework employs a hybrid approach for code retrieval to improve contextual relevance, track inter-file modular dependencies, generate more robust code and ensure consistency with the existing codebase. We benchmark the proposed approach on the Evolutionary Code Benchmark (EvoCodeBench) dataset, a repository-level code generation benchmark, and demonstrate that our method significantly outperforms the baseline approach. These findings suggest that knowledge graph based code generation could advance robust, context-sensitive coding assistance tools. The landscape of software development has fundamentally transformed with the emergence of large language models (LLMs) such as GPT -4 [1] and Claude-3.5-Sonnet,

This paper explores the design of an aspect-based sentiment analysis system using large language models (LLMs) for real-world use. We focus on quadruple opinion extraction -- identifying aspect categories, sentiment polarity, targets, and opinion expressions from text data across different domains and languages. We investigate whether a single fine-tuned model can effectively handle multiple domain-specific taxonomies simultaneously. We demonstrate that a combined multi-domain model achieves performance comparable to specialized single-domain models while reducing operational complexity. We also share lessons learned for handling non-extractive predictions and evaluating various failure modes when developing LLM-based systems for structured prediction tasks.

Lane Keeping Assist (LKA) is widely adopted in modern vehicles, yet its real-world performance remains underexplored due to proprietary systems and limited data access. This paper presents OpenLKA, the first open, large-scale dataset for LKA evaluation and improvement. It includes 400 hours of driving data from 62 production vehicle models, collected through extensive road testing in Tampa, Florida and global contributions from the Comma.ai driving community. The dataset spans a wide range of challenging scenarios, including complex road geometries, degraded lane markings, adverse weather, lighting conditions and surrounding traffic. The dataset is multimodal, comprising: i) full CAN bus streams, decoded using custom reverse-engineered DBC files to extract key LKA events (e.g., system disengagements, lane detection failures); ii) synchronized high-resolution dash-cam video; iii) real-time outputs from Openpilot, providing accurate estimates of road curvature and lane positioning; iv) enhanced scene annotations generated by Vision Language Models, describing lane visibility, pavement quality, weather, lighting, and traffic conditions. By integrating vehicle-internal signals with high-fidelity perception and rich semantic context, OpenLKA provides a comprehensive platform for benchmarking the real-world performance of production LKA systems, identifying safety-critical operational scenarios, and assessing the readiness of current road infrastructure for autonomous driving. The dataset is publicly available at: https://github.com/OpenLKA/OpenLKA.

Task-oriented dialogue (TOD) systems are experiencing a revolution driven by Large Language Models (LLMs), yet the evaluation methodologies for these systems remain insufficient for their growing sophistication. While traditional automatic metrics effectively assessed earlier modular systems, they focus solely on the dialogue level and cannot detect critical intermediate errors that can arise during user-agent interactions. In this paper, we introduce TD-EVAL (Turn and Dialogue-level Evaluation), a two-step evaluation framework that unifies fine-grained turn-level analysis with holistic dialogue-level comparisons. At turn level, we evaluate each response along three TOD-specific dimensions: conversation cohesion, backend knowledge consistency, and policy compliance. Meanwhile, we design TOD Agent Arena that uses pairwise comparisons to provide a measure of dialogue-level quality. Through experiments on MultiWOZ 2.4 and τ-Bench, we demonstrate that TD-EVAL effectively identifies the conversational errors that conventional metrics miss. Furthermore, TD-EVAL exhibits better alignment with human judgments than traditional and LLM-based metrics. These findings demonstrate that TD-EVAL introduces a new paradigm for TOD system evaluation, efficiently assessing both turn and system levels with a plug-and-play framework for future research.