Large Language Models employing extended chain-of-thought (CoT) reasoning often suffer from the overthinking phenomenon, generating excessive and redundant reasoning steps that increase computational costs while potentially degrading performance. While recent work has explored static steering approaches to mitigate this issue, they lack the adaptability to dynamically adjust intervention strength based on real-time reasoning quality. We propose STUPID (Steering Token Usage via PID controller), a novel training-free method that employs a PID controller to dynamically modulate activation steering strength during inference. Our approach combines a chunk-level classifier for detecting redundant reasoning patterns with a PID control mechanism that adaptively adjusts steering intensity based on the predicted redundancy probability. Experimental evaluation on GSM8K demonstrates that STUPID achieves a 6% improvement in accuracy while reducing token usage by 32%, outperforming static steering baselines. Our method provides a principled framework for dynamic reasoning calibration that maintains reasoning quality while significantly improving computational efficiency.

Large Language Models (LLMs) have been extensively evaluated for general summarization tasks as well as medical research assistance, but they have not been specifically evaluated for the task of summarizing real-world evidence (RWE) from structured output of RWE studies. We introduce RWESummary, a proposed addition to the MedHELM framework (Bedi, Cui, Fuentes, Unell et al., 2025) to enable benchmarking of LLMs for this task. RWESummary includes one scenario and three evaluations covering major types of errors observed in summarization of medical research studies and was developed using Atropos Health proprietary data. Additionally, we use RWESummary to compare the performance of different LLMs in our internal RWE summarization tool. At the time of publication, with 13 distinct RWE studies, we found the Gemini 2.5 models performed best overall (both Flash and Pro). We suggest RWESummary as a novel and useful foundation model benchmark for real-world evidence study summarization.

Large Language Models (LLMs) have grown increasingly powerful, yet ensuring their decisions remain transparent and trustworthy requires self-consistency -- no contradictions in their internal reasoning. Our study reveals that even on simple tasks, such as comparing points on a line or a plane, or reasoning in a family tree, all smaller models are highly inconsistent, and even state-of-the-art models like DeepSeek-R1 and GPT-o4-mini are not fully self-consistent. To quantify and mitigate these inconsistencies, we introduce inconsistency metrics and propose two automated methods -- a graph-based and an energy-based approach. While these fixes provide partial improvements, they also highlight the complexity and importance of self-consistency in building more reliable and interpretable AI. The code and data are available at https://github.com/scorpio-nova/llm-self-consistency.

Training large language models (LLMs) on source code significantly enhances their general-purpose reasoning abilities, but the mechanisms underlying this generalisation are poorly understood. In this paper, we propose Programming by Backprop (PBB) as a potential driver of this effect - teaching a model to evaluate a program for inputs by training on its source code alone, without ever seeing I/O examples. To explore this idea, we finetune LLMs on two sets of programs representing simple maths problems and algorithms: one with source code and I/O examples (w/ IO), the other with source code only (w/o IO). We find evidence that LLMs have some ability to evaluate w/o IO programs for inputs in a range of experimental settings, and make several observations. Firstly, PBB works significantly better when programs are provided as code rather than semantically equivalent language descriptions. Secondly, LLMs can produce outputs for w/o IO programs directly, by implicitly evaluating the program within the forward pass, and more reliably when stepping through the program in-context via chain-of-thought. We further show that PBB leads to more robust evaluation of programs across inputs than training on I/O pairs drawn from a distribution that mirrors naturally occurring data. Our findings suggest a mechanism for enhanced reasoning through code training: it allows LLMs to internalise reusable algorithmic abstractions. Significant scope remains for future work to enable LLMs to more effectively learn from symbolic procedures, and progress in this direction opens other avenues like model alignment by training on formal constitutional principles.

In real-world applications, Large Language Models (LLMs) often hallucinate, even in Retrieval-Augmented Generation (RAG) settings, which poses a significant challenge to their deployment. In this paper, we introduce AggTruth, a method for online detection of contextual hallucinations by analyzing the distribution of internal attention scores in the provided context (passage). Specifically, we propose four different variants of the method, each varying in the aggregation technique used to calculate attention scores. Across all LLMs examined, AggTruth demonstrated stable performance in both same-task and cross-task setups, outperforming the current SOTA in multiple scenarios. Furthermore, we conducted an in-depth analysis of feature selection techniques and examined how the number of selected attention heads impacts detection performance, demonstrating that careful selection of heads is essential to achieve optimal results.

This study examines how large language models understand the concept of persuasiveness in public speaking by modifying speech transcripts from PhD candidates in the "Ma These en 180 Secondes" competition, using the 3MT French dataset. Our contributions include a novel methodology and an interpretable textual feature set integrating rhetorical devices and discourse markers. We prompt GPT-4o to enhance or diminish persuasiveness and analyze linguistic shifts between original and generated speech in terms of the new features. Results indicate that GPT-4o applies systematic stylistic modifications rather than optimizing persuasiveness in a human-like manner. Notably, it manipulates emotional lexicon and syntactic structures (such as interrogative and exclamatory clauses) to amplify rhetorical impact.

Continuous chain-of-thought has been shown to be effective in saving reasoning tokens for large language models. By reasoning with continuous latent thought tokens, continuous CoT is able to perform implicit reasoning in a compact manner. However, the sequential dependencies between latent thought tokens spoil parallel training, leading to long training time. In this paper, we propose Parallel Continuous Chain-of-Thought (PCCoT), which performs Jacobi iteration on the latent thought tokens, updating them iteratively in parallel instead of sequentially and thus improving both training and inference efficiency of continuous CoT. Experiments demonstrate that by choosing the proper number of iterations, we are able to achieve comparable or even better performance while saving nearly 50% of the training and inference time. Moreover, PCCoT shows better stability and robustness in the training process. Our code is available at https://github.com/whyNLP/PCCoT.

Detecting harmful content is a crucial task in the landscape of NLP applications for Social Good, with hate speech being one of its most dangerous forms. But what do we mean by hate speech, how can we define it, and how does prompting different definitions of hate speech affect model performance? The contribution of this work is twofold. At the theoretical level, we address the ambiguity surrounding hate speech by collecting and analyzing existing definitions from the literature. We organize these definitions into a taxonomy of 14 Conceptual Elements-building blocks that capture different aspects of hate speech definitions, such as references to the target of hate (individual or groups) or of the potential consequences of it. At the experimental level, we employ the collection of definitions in a systematic zero-shot evaluation of three LLMs, on three hate speech datasets representing different types of data (synthetic, human-in-the-loop, and real-world). We find that choosing different definitions, i.e., definitions with a different degree of specificity in terms of encoded elements, impacts model performance, but this effect is not consistent across all architectures.

The widespread deployment of large language models (LLMs) has raised critical concerns over their vulnerability to jailbreak attacks, i.e., adversarial prompts that bypass alignment mechanisms and elicit harmful or policy-violating outputs. While proprietary models like GPT-4 have undergone extensive evaluation, the robustness of emerging open-source alternatives such as DeepSeek remains largely underexplored, despite their growing adoption in real-world applications. In this paper, we present the first systematic jailbreak evaluation of DeepSeek-series models, comparing them with GPT-3.5 and GPT-4 using the Harm-Bench benchmark. We evaluate seven representative attack strategies across 510 harmful behaviors categorized by both function and semantic domain. Our analysis reveals that DeepSeek's Mixture-of-Experts (MoE) architecture introduces routing sparsity that offers selective robustness against optimization-based attacks such as TAP-T, but leads to significantly higher vulnerability under prompt-based and manually engineered attacks. In contrast, GPT-4 Turbo demonstrates stronger and more consistent safety alignment across diverse behaviors, likely due to its dense Transformer design and reinforcement learning from human feedback. Fine-grained behavioral analysis and case studies further show that DeepSeek often routes adversarial prompts to under-aligned expert modules, resulting in inconsistent refusal behaviors. These findings highlight a fundamental trade-off between architectural efficiency and alignment generalization, emphasizing the need for targeted safety tuning and modular alignment strategies to ensure secure deployment of open-source LLMs.

In the design process of the analog circuit pre-layout phase, device sizing is an important step in determining whether an analog circuit can meet the required performance metrics. Many existing techniques extract the circuit sizing task as a mathematical optimization problem to solve and continuously improve the optimization efficiency from a mathematical perspective. But they ignore the automatic introduction of prior knowledge, fail to achieve effective pruning of the search space, which thereby leads to a considerable compression margin remaining in the search space. To alleviate this problem, we propose a large language model (LLM)-based multi-agent framework for analog circuits' sizing relationships extraction from academic papers. The search space in the sizing process can be effectively pruned based on the sizing relationship extracted by this framework. Eventually, we conducted tests on 3 types of circuits, and the optimization efficiency was improved by $2.32 \sim 26.6 \times$. This work demonstrates that the LLM can effectively prune the search space for analog circuit sizing, providing a new solution for the combination of LLMs and conventional analog circuit design automation methods.