High-order numerical methods enhance Transformer performance in tasks like NLP and CV, but introduce a performance-efficiency trade-off due to increased computational overhead. Our analysis reveals that conventional efficiency techniques, such as distillation, can be detrimental to the performance of these models, exemplified by PCformer. To explore more optimizable ODE-based Transformer architectures, we propose the Iterative Implicit Euler Transformer (IIET), which simplifies high-order methods using an iterative implicit Euler approach. This simplification not only leads to superior performance but also facilitates model compression compared to PCformer. To enhance inference efficiency, we introduce Iteration Influence-Aware Distillation (IIAD). Through a flexible threshold, IIAD allows users to effectively balance the performance-efficiency trade-off. On lm-evaluation-harness, IIET boosts average accuracy by 2.65% over vanilla Transformers and 0.8% over PCformer. Its efficient variant, E-IIET, significantly cuts inference overhead by 55% while retaining 99.4% of the original task accuracy. Moreover, the most efficient IIET variant achieves an average performance gain exceeding 1.6% over vanilla Transformer with comparable speed.

Unsupervised analysis of text corpora is challenging, especially in data-scarce domains where traditional topic models struggle. While these models offer a solution, they typically describe clusters with lists of keywords that require significant manual effort to interpret and often lack semantic coherence. To address this critical interpretability gap, we introduce Recursive Thematic Partitioning (RTP), a novel framework that leverages Large Language Models (LLMs) to interactively build a binary tree. Each node in the tree is a natural language question that semantically partitions the data, resulting in a fully interpretable taxonomy where the logic of each cluster is explicit. Our experiments demonstrate that RTP's question-driven hierarchy is more interpretable than the keyword-based topics from a strong baseline like BERTopic. Furthermore, we establish the quantitative utility of these clusters by showing they serve as powerful features in downstream classification tasks, particularly when the data's underlying themes correlate with the task labels. RTP introduces a new paradigm for data exploration, shifting the focus from statistical pattern discovery to knowledge-driven thematic analysis. Furthermore, we demonstrate that the thematic paths from the RTP tree can serve as structured, controllable prompts for generative models. This transforms our analytical framework into a powerful tool for synthesis, enabling the consistent imitation of specific characteristics discovered in the source corpus.

Recent advances in reinforcement learning (RL) have significantly enhanced the agentic capabilities of large language models (LLMs). In long-term and multi-turn agent tasks, existing approaches driven solely by outcome rewards often suffer from the problem of sparse supervision. To address the challenge, we propose Tree-based Group Relative Policy Optimization (Tree-GRPO), a grouped agent RL method based on tree search, where each tree node represents the complete agent interaction step. By sharing common prefixes, the tree search sampling increases the number of rollouts achievable within a fixed budget of tokens or tool calls. Moreover, we find that the tree-structured trajectory naturally allows the construction of step-wise process supervised signals even using only the outcome reward. Based on this, Tree-GRPO estimates the grouped relative advantages both on intra-tree and inter-tree levels. Through theoretical analysis, we demonstrate that the objective of intra-tree level group relative policy optimization is equivalent to that of step-level direct preference learning. Experiments across 11 datasets and 3 types of QA tasks demonstrate the superiority of the proposed tree-based RL over the chain-based RL method.Figure 1: Comparison of chain-based and tree-based sampling strategies in LLM multi-turn agent RL. The tree structure brings two major advantages: (i) less rollout budget (both on tokens and tool-calls); (ii) higher performance. Reinforcement Learning (RL) has emerged as a pivotal post-training paradigm for Large Language Models (LLMs), catalyzing the development of several frontier models (DeepSeek-AI Team, 2025; Y ang et al., 2025a; OpenAI, 2024). RL-tuned LLMs trained only with outcome rewards acquire complex reasoning abilities and achieve remarkable gains in single-turn tasks, such as mathematical proof and code generation (Team et al., 2025b; Y u et al., 2025; Chu et al., 2025a; Shao et al., 2024; Xin et al., 2024). This suggests that LLMs can learn not only through static imitation, but also by actively interacting with dynamic environments. Guided by this prospect, recent works have extended this RL paradigm to more complex agent settings involving dynamic, multi-turn interactions (Feng et al., 2025b; Singh et al., 2025; Wang et al., 2025b; Qian et al., 2025; Feng et al., Work done during internship at AMAP, Alibaba Group. Right (Ours): Tree search with nodes corresponding to complete agent step.

Modern large language models (LLMs) extend context lengths to millions of tokens, enabling coherent, personalized responses grounded in long conversational histories. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly becomes the bottleneck in resource-constrained environments. An active line of research for reducing memory bottleneck is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting the KV cache after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to failure cases in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40%, maintains near-full KV accuracy under 4-6x compression, and reduces latency/memory by up to 2.4x/3.5x, enabling efficient multi-turn interaction under strict resource limits. Our code is available at https://github.com/apple/ml-epicache.

Adapting billion-parameter language models to a downstream task is still costly, even with parameter-efficient fine-tuning (PEFT). We re-cast task adaptation as output-distribution alignment: the objective is to steer the output distribution toward the task distribution directly during decoding rather than indirectly through weight updates. Building on this view, we introduce Steering Vector Decoding (SVDecode), a lightweight, PEFT-compatible, and theoretically grounded method. We start with a short warm-start fine-tune and extract a task-aware steering vector from the Kullback-Leibler (KL) divergence gradient between the output distribution of the warm-started and pre-trained models. This steering vector is then used to guide the decoding process to steer the model's output distribution towards the task distribution. We theoretically prove that SVDecode is first-order equivalent to the gradient step of full fine-tuning and derive a globally optimal solution for the strength of the steering vector. Across three tasks and nine benchmarks, SVDecode paired with four standard PEFT methods improves multiple-choice accuracy by up to 5 percentage points and open-ended truthfulness by 2 percentage points, with similar gains (1-2 percentage points) on commonsense datasets without adding trainable parameters beyond the PEFT adapter. SVDecode thus offers a lightweight, theoretically grounded path to stronger task adaptation for large language models.

This paper presents ASTREA, the first agentic system executed on flight-heritage hardware (TRL 9) for autonomous spacecraft operations, with on-orbit operation aboard the International Space Station (ISS). Using thermal control as a representative use case, we integrate a resource-constrained Large Language Model (LLM) agent with a reinforcement learning controller in an asynchronous architecture tailored for space-qualified platforms. Ground experiments show that LLM-guided supervision improves thermal stability and reduces violations, confirming the feasibility of combining semantic reasoning with adaptive control under hardware constraints. On-orbit validation aboard the ISS initially faced challenges due to inference latency misaligned with the rapid thermal cycles of Low Earth Orbit (LEO) satellites. Synchronization with the orbit length successfully surpassed the baseline with reduced violations, extended episode durations, and improved CPU utilization. These findings demonstrate the potential for scalable agentic supervision architectures in future autonomous spacecraft.

Large Language Models (LLMs) are widely used in industry but remain prone to hallucinations, limiting their reliability in critical applications. This work addresses hallucination reduction in consumer grievance chatbots built using LLaMA 3.1 8B Instruct, a compact model frequently used in industry. We develop HalluDetect, an LLM-based hallucination detection system that achieves an F1 score of 68.92% outperforming baseline detectors by 22.47%. Benchmarking five hallucination mitigation architectures, we find that out of them, AgentBot minimizes hallucinations to 0.4159 per turn while maintaining the highest token accuracy (96.13%), making it the most effective mitigation strategy. Our findings provide a scalable framework for hallucination mitigation, demonstrating that optimized inference strategies can significantly improve factual accuracy.

Predicting the success of start-up companies, defined as achieving an exit through acquisition or IPO, is a critical problem in entrepreneurship and innovation research. Datasets such as Crunchbase provide both structured information (e.g., funding rounds, industries, investor networks) and unstructured text (e.g., company descriptions), but effectively leveraging this heterogeneous data for prediction remains challenging. Traditional machine learning approaches often rely only on structured features and achieve moderate accuracy, while large language models (LLMs) offer rich reasoning abilities but struggle to adapt directly to domain-specific business data. We present \textbf{CrunchLLM}, a domain-adapted LLM framework for startup success prediction. CrunchLLM integrates structured company attributes with unstructured textual narratives and applies parameter-efficient fine-tuning strategies alongside prompt optimization to specialize foundation models for entrepreneurship data. Our approach achieves accuracy exceeding 80\% on Crunchbase startup success prediction, significantly outperforming traditional classifiers and baseline LLMs. Beyond predictive performance, CrunchLLM provides interpretable reasoning traces that justify its predictions, enhancing transparency and trustworthiness for financial and policy decision makers. This work demonstrates how adapting LLMs with domain-aware fine-tuning and structured--unstructured data fusion can advance predictive modeling of entrepreneurial outcomes. CrunchLLM contributes a methodological framework and a practical tool for data-driven decision making in venture capital and innovation policy.

Smart contracts automate the management of high-value assets, where vulnerabilities can lead to catastrophic financial losses. This challenge is amplified in Large Language Models (LLMs) by two interconnected failures: they operate as unauditable "black boxes" lacking a transparent reasoning process, and consequently, generate code riddled with critical security vulnerabilities. To address both issues, we propose SmartCoder-R1 (based on Qwen2.5-Coder-7B), a novel framework for secure and explainable smart contract generation. It begins with Continual Pre-training (CPT) to specialize the model. We then apply Long Chain-of-Thought Supervised Fine-Tuning (L-CoT SFT) on 7,998 expert-validated reasoning-and-code samples to train the model to emulate human security analysis. Finally, to directly mitigate vulnerabilities, we employ Security-Aware Group Relative Policy Optimization (S-GRPO), a reinforcement learning phase that refines the generation policy by optimizing a weighted reward signal for compilation success, security compliance, and format correctness. Evaluated against 17 baselines on a benchmark of 756 real-world functions, SmartCoder-R1 establishes a new state of the art, achieving top performance across five key metrics: a ComPass of 87.70%, a VulRate of 8.60%, a SafeAval of 80.16%, a FuncRate of 53.84%, and a FullRate of 50.53%. This FullRate marks a 45.79% relative improvement over the strongest baseline, DeepSeek-R1. Crucially, its generated reasoning also excels in human evaluations, achieving high-quality ratings for Functionality (82.7%), Security (85.3%), and Clarity (90.7%).

We propose FSPO (Fair Sequence Policy Optimization), a sequence-level reinforcement learning method for LLMs that enforces length-fair clipping on the importance-sampling (IS) weight. We study RL methods with sequence-level IS and identify a mismatch when PPO/GRPO-style clipping is transplanted to sequences: a fixed clip range systematically reweights short vs. long responses, distorting the optimization direction. FSPO introduces a simple remedy: we clip the sequence log-IS ratio with a band that scales as $\sqrt{L}$. Theoretically, we formalize length fairness via a Length Reweighting Error (LRE) and prove that small LRE yields a cosine directional guarantee between the clipped and true updates. Empirically, FSPO flattens clip rates across length bins, stabilizes training, and outperforms baselines across model sizes and evaluation datasets, with the largest gains on the Qwen3-8B-Base model.