--Register Transfer Level(RTL) code optimization is crucial for achieving high performance and low power consumption in digital circuit design. However, traditional optimization methods often rely on manual tuning and heuristics, which can be time-consuming and error-prone. Recent studies proposed to leverage Large Language Models(LLMs) to assist in RTL code optimization. LLMs can generate optimized code snippets based on natural language descriptions, potentially speeding up the optimization process. However, existing approaches have not thoroughly evaluated the effectiveness of LLM-Based code optimization methods for RTL code with complex timing logic. T o address this gap, we conducted a comprehensive empirical investigation to assess the capability of LLM-Based RTL code optimization methods in handling RTL code with complex timing logic. In this study, we first propose a new benchmark for RTL optimization evaluation. It comprises four subsets, each corresponding to a specific area of RTL code optimization, logic operation optimization, data path optimization, timing control flow optimization, and clock domain optimization . Then we introduce a method based on metamorphosis to systematically evaluate the effectiveness of LLM-Based RTL code optimization methods. Our key insight is that the optimization effectiveness should remain consistent for semantically equivalent but more complex code. After intensive experiments, we revealed several key findings. This is primarily attributed to the challenges LLMs face in understanding timing logic in RTL code. Based on these findings, we provide insights for further research in leveraging LLMs for RTL code optimization. Optimizing Register Transfer Level (RTL) code is a critical step in the early stages of circuit design [1]-[4].

When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any analogous reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models.

While modern dialogue systems heavily rely on large language models (LLMs), their implementation often goes beyond pure LLM interaction. Developers integrate multiple LLMs, external tools, and databases. Therefore, assessment of the underlying LLM alone does not suffice, and the dialogue systems must be tested and evaluated as a whole. However, this remains a major challenge. With most previous work focusing on turn-level analysis, less attention has been paid to integrated dialogue-level quality assurance. To address this, we present ChatChecker, a framework for automated evaluation and testing of complex dialogue systems. ChatChecker uses LLMs to simulate diverse user interactions, identify dialogue breakdowns, and evaluate quality. Compared to previous approaches, our design reduces setup effort and is generalizable, as it does not require reference dialogues and is decoupled from the implementation of the target dialogue system. We improve breakdown detection performance over a prior LLM-based approach by including an error taxonomy in the prompt. Additionally, we propose a novel non-cooperative user simulator based on challenging personas that uncovers weaknesses in target dialogue systems more effectively. Through this, ChatChecker contributes to thorough and scalable testing. This enables both researchers and practitioners to accelerate the development of robust dialogue systems.

Performance is achieved by modeling natural language as reasoning trees measured by both length and depth instead of linear sequences. The reasoning trees consist of tasks with thoughts, recursive sub-tasks, and conclusions based on the concept we proposed in (Schroeder et al., 2025). During generation, we maintain a working memory that retains only the key/value states of the most relevant context tokens, selected by a rule-based subtask-pruning mechanism, enabling reuse of positional embeddings and GPU memory pages throughout reasoning. Experimental results show that our system sustains high inference throughput, even when manipulating up to 90% of the KV cache in GPU memory. It also delivers accurate reasoning on mathematical tasks and handles information retrieval challenges that require long-horizon reasoning and multi-hop tool use. Large language models (LLMs) have emerged as versatile foundations for a wide range of AI applications, especially agents which handle complicated tasks including multi-hop reasoning and tool use. Their ability to generalize across various tasks with minimal fine-tuning has driven rapid innovation and broad adoption (Brown et al., 2020). However, the fundamental objective of language modeling, to generate unstructured token sequences (Bengio et al., 2003), imposes strict context window limits and makes fine-grained control over internal state difficult. As a result, these inherent constraints pose significant challenges for all state-of-the-art LLMs, notably their inability to maintain long-horizon reasoning trajectories and coordinate complex workflows, which hinders the development of robust, memory-intensive applications. Neural networks generate natural language as a linear sequence. Recurrent neural networks (RNNs) (Mikolov et al., 2010; Luong et al., 2015; Gu & Dao, 2023) and Transformers (V aswani et al., 2017; Y ang et al., 2023) are constrained by token limits, hidden state sizes, and GPU-memory capacities.

Structured decoding enables large language models (LLMs) to generate outputs in formats required by downstream systems, such as HTML or JSON. However, existing methods suffer from efficiency bottlenecks due to grammar compilation, state tracking, and mask creation. We observe that many real-world tasks embed strong prior knowledge about output structure. Leveraging this, we propose a decomposition of constraints into static and dynamic components -- precompiling static structures offline and instantiating dynamic arguments at runtime using grammar snippets. Instead of relying on pushdown automata, we employ a compositional set of operators to model regular formats, achieving lower transition latency. We introduce wgrammar, a lightweight decoding engine that integrates domain-aware simplification, constraint decomposition, and mask caching, achieving up to 250x speedup over existing systems. wgrammar's source code is publicly available at https://github.com/wrran/wgrammar.

Vision-language models (VLMs) have been widely adopted in robotics to enable autonomous planning. However, grounding VLMs, originally trained on internet data, to diverse real-world robots remains a challenge. This paper presents ExpTeach, a framework that grounds VLMs to physical robots by building a self-generated memory of real-world experiences. In ExpTeach, the VLM autonomously plans actions, verifies outcomes, reflects on failures, and adapts robot behaviors in a closed loop. The self-generated experiences during this process are then summarized into a long-term memory, enabling retrieval of learned knowledge to guide future tasks via retrieval-augmented generation (RAG). Additionally, ExpTeach enhances the spatial understanding of VLMs with an on-demand image annotation module. In experiments, we show that reflection improves success rates from 36% to 84% on four challenging robotic tasks and observe the emergence of intelligent object interactions, including creative tool use. Across extensive tests on 12 real-world scenarios (including eight unseen ones), we find that grounding with long-term memory boosts single-trial success rates from 22% to 80%, demonstrating the effectiveness and generalizability of ExpTeach.

In-Context Learning has shown great potential for aligning Large Language Models (LLMs) with human values, helping reduce harmful outputs and accommodate diverse preferences without costly post-training, known as In-Context Alignment (ICA). However, LLMs' comprehension of input prompts remains agnostic, limiting ICA's ability to address value tensions--human values are inherently pluralistic, often imposing conflicting demands, e.g., stimulation vs. tradition. Current ICA methods therefore face the Instruction Bottleneck challenge, where LLMs struggle to reconcile multiple intended values within a single prompt, leading to incomplete or biased alignment. To address this, we propose PICACO, a novel pluralistic ICA method. Without fine-tuning, PICACO optimizes a meta-instruction that navigates multiple values to better elicit LLMs' understanding of them and improve their alignment. This is achieved by maximizing the total correlation between specified values and LLM responses, theoretically reinforcing value correlation while reducing distractive noise, resulting in effective value instructions. Extensive experiments on five value sets show that PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves a better balance across up to 8 distinct values.

Transformers and large language models (LLMs), powered by the attention mechanism, have transformed numerous AI applications, driving the need for specialized hardware accelerators. A major challenge in these accelerators is efficiently detecting errors caused by random hardware faults. Traditional algorithm-based fault tolerance (ABFT) techniques verify individual matrix multiplications but fall short in handling the full attention mechanism, particularly due to intermediate softmax normalization. This work proposes Flash-ABFT, a novel method that computes an online checksum across the entire three-matrix product of query, key and value matrices, of an attention layer, including the softmax operation, with a single check. This approach significantly reduces overhead by eliminating redundant checks while maintaining high fault-detection accuracy. Experimental results demonstrate that Flash-ABFT incurs only 5.3% hardware area overhead and less than 1.9% energy overhead, making it a cost-effective and robust solution for error detection in attention accelerators.

Generative, explainable, and flexible recommender systems, derived using Large Language Models (LLM) are promising and poorly adapted to the cold-start user situation, where there is little to no history of interaction. The current solutions i.e. supervised fine-tuning and collaborative filtering are dense-user-item focused and would be expensive to maintain and update. This paper introduces a meta-learning framework, that can be used to perform parameter-efficient prompt-tuning, to effectively personalize LLM-based recommender systems quickly at cold-start. The model learns soft prompt embeddings with first-order (Reptile) and second-order (MAML) optimization by treating each of the users as the tasks. As augmentations to the input tokens, these learnable vectors are the differentiable control variables that represent user behavioral priors. The prompts are meta-optimized through episodic sampling, inner-loop adaptation, and outer-loop generalization. On MovieLens-1M, Amazon Reviews, and Recbole, we can see that our adaptive model outperforms strong baselines in NDCG@10, HR@10, and MRR, and it runs in real-time (i.e., below 300 ms) on consumer GPUs. Zero-history personalization is also supported by this scalable solution, and its 275 ms rate of adaptation allows successful real-time risk profiling of financial systems by shortening detection latency and improving payment network stability. Crucially, the 275 ms adaptation capability can enable real-time risk profiling for financial institutions, reducing systemic vulnerability detection latency significantly versus traditional compliance checks. By preventing contagion in payment networks (e.g., Fedwire), the framework strengthens national financial infrastructure resilience.

This study explores the potential of phonological reasoning within text-based large language models (LLMs). Utilizing the PhonologyBench benchmark, we assess tasks like rhyme word generation, g2p conversion, and syllable counting. Our evaluations across 12 LLMs reveal that while few-shot learning offers inconsistent gains, the introduction of a novel Pedagogically-motivated Participatory Chain-of-Thought (P-CoT) prompt, which is anchored in educational theories like scaffolding and discovery learning, consistently enhances performance. This method leverages structured guidance to activate latent phonological abilities, achieving up to 52% improvement and even surpassing human baselines in certain tasks. Future work could aim to optimize P-CoT prompts for specific models or explore their application across different linguistic domains.