Programming with a coding assistant is a fundamentally interactive process, yet existing static benchmarks fail to capture key features of model-user collaboration. We introduce an interactive evaluation pipeline to examine how LLMs incorporate different types of feedback in a collaborative setting, in which we obfuscate the input of static coding benchmarks so that the code model must interact with a simulated user. Across 10 models and 3 datasets, the relative rankings of models often permute greatly between static and interactive settings, despite models being fairly robust to feedback that contains errors. We also observe that similarly effective feedback types differ in terms of how models respond to higher- vs. lower-quality feedback. Moreover, feedback type impacts the degree to which the models make aesthetic or behavioral edits to their output. Our work aims to “re-evaluate” model coding capabilities through an interactive lens toward bridging the gap between existing evaluations and real-world usage.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
In inference-time scaling, Chain-of-Thought (CoT) plays a crucial role in enabling large language models (LLMs) to exhibit reasoning capabilities. However, in many scenarios, high-quality CoT data is scarce or even unavailable. In such cases, STaR-like methods can help LLMs synthesize CoT based on user queries and response, but they inevitably suffer from the risk of compounding errors. In this work, we tackle an even more challenging scenario: tool learning in the absence of user queries. We design a data scaling method using back-translation, which establishes an inference cycle to synthesize both user queries and CoT data. To reudce the compounding error of inference time, we introduce two rule-based verifiers to assess the validity of the synthesized CoT data. In particular, the Cycle Verifier facilitates performance improvement by continuously accumulating new data over multiple iterations. Our approach achieves a 75.4% pass rate and a 79.6% win rate using small models (7B) in StableToolBench. Notably, these results are obtained exclusively from self-synthesized high-quality data, without relying on external supervision or expert trajectories for warm-up.
Natural language interfaces for NoSQL databases are increasingly vital in the big data era, enabling users to interact with complex, unstructured data without deep technical expertise. However, most recent advancements focus on English, leaving a gap for multilingual support. This paper introduces MultiTEND, the first and largest multilingual benchmark for natural language to NoSQL query generation, covering six languages: English, German, French, Russian, Japanese and Mandarin Chinese.Using MultiTEND, we analyze challenges in translating natural language to NoSQL queries across diverse linguistic structures, including lexical and syntactic differences. Experiments show that performance accuracy in both English and non-English settings remains relatively low, with a 4%-6% gap across scenarios like fine-tuned SLM, zero-shot LLM, and RAG for LLM.To address the aforementioned challenges, we introduce MultiLink, a novel framework that bridges the multilingual input to NoSQL query generation gap through a Parallel Linking Process. It breaks down the task into multiple steps, integrating parallel multilingual processing, Chain-of-Thought (CoT) reasoning, and Retrieval-Augmented Generation (RAG) to tackle lexical and structural challenges inherent in multilingual NoSQL generation. MultiLink shows enhancements in all metrics for every language against the top baseline, boosting execution accuracy by about 15% for English and averaging a 10% improvement for non-English languages.
Small large language models (sLLMs) offer the advantage of being lightweight and efficient, which makes them suitable for resource-constrained environments. However, sLLMs often struggle to maintain topic consistency in task-oriented dialogue systems, which is critical for scenarios such as service chatbots. Specifically, it is important to ensure that the model denies off-topic or malicious inputs and adheres to its intended functionality so as to prevent potential misuse and uphold reliability. Towards this, existing activation engineering approaches have been proposed to manipulate internal activations during inference. While these methods are effective in certain scenarios, our preliminary experiments reveal their limitations in ensuring topic adherence. Therefore, to address this, we propose a novel approach termed Entropy-scaled Steering vectors for Topic Maintenance (EnSToM). EnSToM dynamically adjusts the steering intensity based on input uncertainty, which allows the model to handle off-topic distractors effectively while preserving on-topic accuracy. Our experiments demonstrate that EnSToM achieves significant performance gain with a relatively small data size compared to fine-tuning approaches. By improving topic adherence without compromising efficiency, our approach provides a robust solution for enhancing sLLM-based dialogue systems.
Safeguarding RAG Pipelines with GMTP: A Gradient-based Masked Token Probability Method for Poisoned Document Detection
PDF ↗Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by providing external knowledge for accurate and up-to-date responses. However, this reliance on external sources exposes a security risk; attackers can inject poisoned documents into the knowledge base to steer the generation process toward harmful or misleading outputs. In this paper, we propose Gradient-based Masked Token Probability (GMTP), a novel defense method to detect and filter out adversarially crafted documents. Specifically, GMTP identifies high-impact tokens by examining gradients of the retriever’s similarity function. These key tokens are then masked, and their probabilities are checked via a Masked Language Model (MLM). Since injected tokens typically exhibit markedly low masked-token probabilities, this enables GMTP to easily detect malicious documents and achieve high-precision filtering. Experiments demonstrate that GMTP is able to eliminate over 90% of poisoned content while retaining relevant documents, thus maintaining robust retrieval and generation performance across diverse datasets and adversarial settings.
Real-world applications of KBQA require models to detect different types of unanswerable questions with a limited volume of in-domain labeled training data. We propose the novel task of few-shot transfer for KBQA with unanswerable questions. The state-of-the-art KBQA few-shot transfer model (FuSIC-KBQA) uses an iterative repair strategy that assumes that all questions are answerable. As a remedy, we present FUn-FuSIC – a novel solution for our task that extends FuSIC-KBQA with Feedback for Unanswerability (FUn), which is an iterative repair strategy for answerable as well as unanswerable questions. FUn uses feedback from a suite of strong and weak verifiers, and an adaptation of self-consistency for unanswerability for assessing answerability of questions. Our experiments show that FUn-FuSIC significantly outperforms suitable adaptations of multiple LLM-based and supervised SoTA models on our task, while establishing a new SoTA performance for answerable few-shot transfer as well. We have made datasets and other resources publicly available at https://github.com/dair-iitd/funfusic/.
Entrospect: Information-Theoretic Self-Reflection Elicits Better Response Refinement of Small Language Models
PDF ↗Self-reflection helps de-hallucinate Large Language Models (LLMs). However, the effectiveness of self-reflection remains insufficiently validated in the context of Small Language Models (SLMs), which exhibit limited semantic capacities. In particular, we demonstrate that the conventional self-reflection paradigm, such as Self-Refine, fails to deliver robust response refinement for models with parameter sizes of 10 billion or smaller, even when compared to generations elicited through Chain-of-Thought (CoT) prompting. To improve SLMs’ self-reflection, we redesign Self-Refine and introduce Entrospect (ENTROpy-aware IntroSPECTion), an information-theoretic framework based on prompt engineering.We evaluated Entrospect using accuracy and average time consumption metrics to comprehensively assess its precision and computational efficiency. Experiments conducted across four distinct SLMs and four baseline methods demonstrate that Entrospect achieves state-of-the-art performance on validation tasks. Notably, under identical model and data settings, Entrospect delivers a remarkable improvement of up to 36.2 in reasoning accuracy while enhancing computational efficiency by as much as 10 times compared to its predecessor, Self-Refine.
Knowledge editing enables efficient updates to Large Language Models (LLMs) by modifying specific knowledge without full-model retraining. Among knowledge editing approaches, in-context editing (ICE) stands out for its ability to inject knowledge without modifying the model’s parameters. However, existing ICE approaches directly edit model context without isolating target knowledge from the reasoning path of model inference, resulting in unreliable and low-quality outputs, particularly in multi-hop tasks. To investigate this issue, we analyze the interaction between reasoning path planning and knowledge injection, showing that the reasoning ability of a LLM is usually coupled with its original knowledge, and directly replacing old knowledge with new one could simultaneously hurt the LLM’s performance in task reasoning. Based on these findings, we propose DecKER, a novel ICE framework that separates model reasoning from knowledge editing. Extensive experiments show that DecKER significantly improves multi-hop reasoning performance by mitigating knowledge conflicts and preserving reasoning integrity.
Citation context analysis (CCA) is a field of research studying the role and purpose of citation in scientific discourse. While most of the efforts in CCA have been focused on elaborate characterization schemata to assign function or intent labels to individual citations, the citation context as the basis for such a classification has received rather limited attention. This relative neglect, however, has led to the prevalence of vague definitions and restrictive assumptions, limiting the citation context in its expressiveness. It is a common practice, for example, to restrict the context to the citing sentence. While this simple context conceptualization might be sufficient to assign intent or function classes, it fails to cover the rich information of scientific discourse. To address this concern, we analyze the context conceptualizations of previous works and, to our knowledge, construct the first comprehensive context definition based on the semantic properties of the citing text. To evaluate this definition, we construct and publish the FineCite corpus containing 1,056 manually annotated citation contexts. Our experiments on established CCA benchmarks demonstrate the effectiveness of our fine-grained context definition, showing improvements of up to 25% compared to state-of-the-art approaches. We make our code and data publicly available at https://github.com/lab-paper-code/FineCite.
Tool learning aims to augment large language models (LLMs) with diverse tools, enabling them to act as agents for solving practical tasks. Due to the limited context length of tool-using LLMs, adopting information retrieval (IR) models to select useful tools from large toolsets is a critical initial step. However, the performance of IR models in tool retrieval tasks remains underexplored and unclear. Most tool-use benchmarks simplify this step by manually pre-annotating a small set of relevant tools for each task, which is far from the real-world scenarios. In this paper, we propose ToolRet, a heterogeneous tool retrieval benchmark comprising 7.6k diverse retrieval tasks, and a corpus of 43k tools, collected from existing datasets. We benchmark six types of models on ToolRet. Surprisingly, even the models with strong performance in conventional IR benchmarks, exhibit poor performance on ToolRet. This low retrieval quality degrades the task pass rate of tool-use LLMs. As a further step, we contribute a large-scale training dataset with over 200k instances, which substantially optimizes the tool retrieval ability of IR models.
“You are Beautiful, Body Image Stereotypes are Ugly!” BIStereo: A Benchmark to Measure Body Image Stereotypes in Language Models
PDF ↗While a few high-quality bias benchmark datasets exist to address stereotypes in Language Models (LMs), a notable lack of focus remains on body image stereotypes. To bridge this gap, we propose \textbf{BIStereo}, a suite to uncover LMs’ biases towards people of certain physical appearance characteristics, namely, \textit{skin complexion, body shape, height, attire,} and a \textit{miscellaneous category} including \textit{hair texture, eye color, and more}. Our dataset comprises 40k sentence pairs designed to assess LMs’ biased preference for certain body types. We further include 60k premise-hypothesis pairs designed to comprehensively assess LMs’ preference for fair skin tone. Additionally, we curate 553 tuples consisting of a \textit{body image descriptor, gender, and a stereotypical attribute}, validated by a diverse pool of annotators for physical appearance stereotypes.We propose a metric, \textbf{TriSentBias}, that captures the biased preferences of LMs towards a certain body type over others. Using \textbf{BIStereo}, we assess the presence of body image biases in ten different language models, revealing significant biases in models Muril, XLMR, Llama3, and Gemma. We further evaluate the LMs through downstream NLI and Analogy tasks.Our NLI experiments highlight notable patterns in the LMs that align with the well-documented cognitive bias in humans known as \textbf{\textit{the Halo Effect}}.
Vision-language models (VLMs) integrate textual and visual information, enabling the model to process visual inputs and leverage visual information to generate predictions. Such models are demanding for tasks such as visual question answering, image captioning, and visual grounding. However, some recent work found that VLMs often rely heavily on textual information, ignoring visual information, but are still able to achieve competitive performance in vision-language (VL) tasks. This survey reviews modality collapse analysis work to provide insights into the reason for this unintended behavior. It also reviews probing studies for fine-grained vision-language understanding, presenting current findings on information encoded in VL representations and highlighting potential directions for future research.
We present an interesting application of narrative understanding in the clinical assessment of aphasia, where story retelling tasks are used to evaluate a patient’s communication abilities. This clinical setting provides a framework to help operationalize narrative discourse analysis and an application-focused evaluation method for narrative understanding systems. In particular, we highlight the use of main concepts (MCs)—a list of statements that capture a story’s gist—for aphasic discourse analysis. We then propose automatically generating MCs from novel stories, which experts can edit manually, thus enabling wider adaptation of current assessment tools. We further develop a prompt ensemble method using large language models (LLMs) to automatically generate MCs for a novel story. We evaluate our method on an existing narrative summarization dataset to establish its intrinsic validity. We further apply it to a set of stories that have been annotated with MCs through extensive analysis of retells from non-aphasic and aphasic participants (Kurland et al., 2021, 2025). Our results show that our proposed method can generate most of the gold-standard MCs for stories from this dataset. Finally, we release this dataset of stories with annotated MCs to spur more research in this area.
While personalized recommendations are often desired by users, it can be difficult in practice to distinguish cases of bias from cases of personalization: we find that models generate racially stereotypical recommendations regardless of whether the user revealed their identity intentionally through explicit indications or unintentionally through implicit cues. We demonstrate that when people use large language models (LLMs) to generate recommendations, the LLMs produce responses that reflect both what the user wants and who the user is. We argue that chatbots ought to transparently indicate when recommendations are influenced by a user’s revealed identity characteristics, but observe that they currently fail to do so. Our experiments show that even though a user’s revealed identity significantly influences model recommendations (p < 0.001), model responses obfuscate this fact in response to user queries. This bias and lack of transparency occurs consistently across multiple popular consumer LLMs and for four American racial groups.
A NOTAM or NOtice To AirMen is a crucial notice for different aviation stakeholders, particularly flight crews. It delivers essential notifications about abnormal conditions of Aviation System components such as changes to facilities, hazards, service, procedure that are not known far enough in advance to be publicized through other means. NOTAM messages are short, contain acronyms, and look cryptic in most of the cases. Writing and understanding these messages put heavy cognitive load on its end users. In this work, we take up the task of translating NOTAMs into English natural language using LLMs. Since NOTAMs do not adhere to English grammar rules and have their own decoding rules, large language models (LLMs) cannot translate them without effective prompting. In this paper, we develop a framework to come up with effective prompts to achieve the translations. Our approach uses context-aware semantic prompting techniques, paired with domain-specific rules, to improve the accuracy and clarity of translations. The framework is evaluated using comprehensive experiments (6 LLMs of varying sizes, and with 5 different prompting setups for each) and eight evaluation metrics measuring different aspects of the translation. The results demonstrate that our methodology can produce clear translations that accurately convey the information contained in NOTAMs.
Despite the widespread use of English as a Second or Foreign Language (ESFL), developing syntactico-semantic representations for it is limited — the irregularities in ESFL complicate systematic composition and subsequently the derivation of its semantics.This paper draws on constructivism and proposes a novel Synchronous Hyperedge Replacement Grammar (SHRG)-based constructivist approach to address the challenges. By using constructions as fundamental units, this approach not only accommodates both the idiosyncrasies and the compositional nature of ESFL, but also bridges the gap between literal cues and intended meaning.The feasibility of this constructivist approach is demonstrated using real ESFL data, resulting in a gold-standard, medium-sized syntactico-semantic bank that covers a wide range of ESFL phenomena.
Process Reward Models (PRMs) have demonstrated promising results in mathematical reasoning, but existing process annotation approaches, whether through human annotations or Monte Carlo simulations, remain computationally expensive. In this paper, we introduce Step COmpression for Process Estimation (SCOPE), a novel compression-based approach that significantly reduces annotation costs. We first translate natural language reasoning steps into code and normalize them through Abstract Syntax Tree, then merge equivalent steps to construct a prefix tree. Unlike simulation-based methods that waste numerous samples on estimation, SCOPE leverages a compression-based prefix tree where each root-to-leaf path serves as a training sample, reducing the complexity from O(NMK) to O(N) We construct a large-scale dataset containing 509K samples with only 5% of the computational resources required by previous methods. Empirical results demonstrate that PRMs trained on our dataset consistently outperform existing automated annotation approaches on both Best-of-N strategy and ProcessBench.
Do Emotions Really Affect Argument Convincingness? A Dynamic Approach with LLM-based Manipulation Checks
PDF ↗Emotions have been shown to play a role in argument convincingness, yet this aspect is underexplored in the natural language processing (NLP) community. Unlike prior studies that use static analyses, focus on a single text domain or language, or treat emotion as just one of many factors, we introduce a dynamic framework inspired by manipulation checks commonly used in psychology and social science; leveraging LLM-based manipulation checks, this framework examines the extent to which perceived emotional intensity influences perceived convincingness. Through human evaluation of arguments across different languages, text domains, and topics, we find that in over half of cases, human judgments of convincingness remain unchanged despite variations in perceived emotional intensity; when emotions do have an impact, they more often enhance rather than weaken convincingness.We further analyze whether 11 LLMs behave like humans in the same scenario, finding that while LLMs generally mirror human patterns,they struggle to capture nuanced emotional effects in individual judgments.
Full-Step-DPO: Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning
PDF ↗Direct Preference Optimization (DPO) often struggles with long-chain mathematical reasoning. Existing approaches, such as Step-DPO, typically improve this by focusing on the first erroneous step in the reasoning chain. However, they overlook all other steps and rely heavily on humans or GPT-4 to identify erroneous steps. To address these issues, we propose Full-Step-DPO, a novel DPO framework tailored for mathematical reasoning. Instead of optimizing only the first erroneous step, it leverages step-wise rewards from the entire reasoning chain. This is achieved by training a self-supervised process reward model, which automatically scores each step, providing rewards while avoiding reliance on external signals. Furthermore, we introduce a novel step-wise DPO loss, which dynamically updates gradients based on these step-wise rewards. This endows stronger reasoning capabilities to language models. Extensive evaluations on both in-domain and out-of-domain mathematical reasoning benchmarks across various base language models, demonstrate that Full-Step-DPO achieves superior performance compared to state-of-the-art baselines.
Burn After Reading: Do Multimodal Large Language Models Truly Capture Order of Events in Image Sequences?
PDF ↗This paper introduces the TempVS benchmark, which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models (MLLMs) in image sequences. TempVS consists of three main tests (i.e., event relation inference, sentence ordering and image ordering), each accompanied with a basic grounding test. TempVS requires MLLMs to rely on both visual and linguistic modalities to understand the temporal order of events. We evaluate 38 state-of-the-art MLLMs, demonstrating that models struggle to solve TempVS, with a substantial performance gap compared to human capabilities. We also provide fine-grained insights that suggest promising directions for future research. Our TempVS benchmark data and code are available at https://github.com/yjsong22/TempVS.