EACL 2026 论文列表

The miscalibration of Large Reasoning Models (LRMs) undermines their reliability in high-stakes domains, necessitating methods to accurately estimate the confidence of their long-form, multi-step outputs. To address this gap, we introduce the Reasoning Model Confidence estimation Benchmark (RMCB), a public resource of 347,496 reasoning traces from six popular LRMs across different architectural families. The benchmark is constructed from a diverse suite of datasets spanning high-stakes domains, including clinical, financial, legal, and mathematical reasoning, alongside complex general reasoning benchmarks, with correctness annotations provided for all samples. Using RMCB, we conduct a large-scale empirical evaluation of over ten distinct representation-based methods, spanning sequential, graph-based, and text-based architectures. Our central finding is a persistent trade-off between discrimination (AUROC) and calibration (ECE): text-based encoders achieve the best AUROC (0.672), while structurally-aware models yield the best ECE (0.148), with no single method dominating both. Furthermore, we find that increased architectural complexity does not reliably outperform simpler sequential baselines, suggesting a performance ceiling for methods relying solely on chunk-level hidden states. This work provides the most comprehensive benchmark for this task to date, establishing rigorous baselines and demonstrating the limitations of current representation-based paradigms.

Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

Workflows are a fundamental component of automation in enterprise platforms, enabling the orchestration of tasks, data processing, and system integrations. Despite being widely used, building workflows can be complex, often requiring manual configuration through low-code platforms or visual programming tools. To simplify this process, we explore the use of generative foundation models, particularly vision-language models (VLMs), to automatically generate structured workflows from visual inputs. Translating hand-drawn sketches or computer-generated diagrams into executable workflows is challenging due to the ambiguity of free-form drawings, variations in diagram styles, and the difficulty of inferring execution logic from visual elements. To address this, we introduce StarFlow, a framework for generating structured workflow outputs from sketches using vision-language models. We curate a diverse dataset of workflow diagrams – including synthetic, manually annotated, and real-world samples – to enable robust training and evaluation. We finetune and benchmark multiple vision-language models, conducting a series of ablation studies to analyze the strengths and limitations of our approach. Our results show that finetuning significantly enhances structured workflow generation, outperforming large vision-language models on this task.

Adapting large language models to full document translation remains challenging due to the difficulty of capturing long-range dependencies and preserving discourse coherence throughout extended texts. While recent agentic machine translation systems mitigate context window constraints through multi-agent orchestration and persistent memory, they require substantial computational resources and are sensitive to memory retrieval strategies. We introduce TransGraph, a discourse-guided framework that explicitly models inter-chunk relationships through structured discourse graphs and selectively conditions each translation segment on relevant graph neighbourhoods rather than relying on sequential or exhaustive context. Across three document-level MT benchmarks spanning six languages and diverse domains, TransGraph consistently surpasses strong baselines in translation quality and terminology consistency while incurring significantly lower token overhead.

Geocoding is the task of linking a location reference to an actual geographic location and is essential for many downstream analyses of unstructured text. In this paper, we explore the challenging setting of geocoding compositional location references. Building on recent work demonstrating LLMs’ abilities to reason over geospatial data, we evaluate LLMs’ geospatial knowledge versus reasoning skills relevant to our task. Based on these insights, we propose an LLM-based strategy for geocoding compositional location references. We show that our approach improves performance for the task and that a relatively small fine-tuned LLM can achieve comparable performance with much larger off-the-shelf models.

Detecting mental health crisis situations such as suicide ideation, rape, domestic violence, child abuse, and sexual harassment is a critical yet underexplored challenge for language models. When such situations arise during user–model interactions, models must reliably flag them, as failure to do so can have serious consequences. In this work, we introduce CRADLE BENCH, a benchmark for multi-faceted crisis detection. Unlike previous efforts that focus on a limited set of crisis types, our benchmark covers seven types defined in line with clinical standards and is the first to incorporate temporal labels. Our benchmark provides 600 clinician-annotated evaluation examples and 420 development examples, together with a training corpus of around 4K examples automatically labeled using a majority-vote ensemble of multiple language models, which significantly outperforms single-model annotation. We further fine-tune six crisis detection models on subsets defined by consensus and unanimous ensemble agreement, providing complementary models trained under different agreement criteria.Content warning: This paper discusses sensitive topics such as suicide ideation, self-harm, rape, domestic violence, and child abuse.

Large language models (LLMs) struggle with ex-ante reasoning—making inferences or predictions without access to future information. Even under explicit temporal cutoffs, they often rely on internalized post-cutoff knowledge. To systematically evaluate this issue, we introduce a benchmark that assesses LLMs’ ex-ante inference ability across four tasks: stock prediction, question answering, Wikipedia event generation, and scientific publication generation. We quantify temporal leakage using a leakage rate metric, which measures models’ reliance on future information beyond cutoff timestamps, and a quality measure that evaluates task performance. Experimental results show that LLMs frequently violate temporal constraints across tasks, revealing persistent challenges in ex-ante reasoning. Our benchmark serves as a rigorous testbed for studying temporal reasoning in time-sensitive contexts and provides complete datasets, results, and evaluation resources to support future research on improving temporal consistency in modern LLMs.

Understanding narratives requires identifying which events are most salient for a story’s progression. We present a contrastive learning framework for modeling narrative salience that learns story embeddings from narrative twins: stories that share the same plot but differ in surface form. Our model is trained to distinguish a story from both its narrative twin and a distractor with similar surface features but different plot. Using the resulting embeddings, we evaluate four narratologically motivated operations for inferring salience (deletion, shifting, disruption, and summarization). Experiments on short narratives from the ROCStories corpus and longer Wikipedia plot summaries show that contrastively learned story embeddings outperform a masked-language-model baseline, and that summarization is the most reliable operation for identifying salient sentences. If narrative twins are not available, random dropout can be used to generate the twins from a single story. Effective distractors can be obtained either by prompting LLMs or, in long-form narratives, by using different parts of the same story.

We introduce TDFlow, a novel test-driven agentic workflow that frames repository-scale software engineering as a test-resolution task, specifically designed to solve human-written tests. Given a set of tests, TDFlow repeatedly proposes, revises, and debugs repository-scale patches using precisely engineered sub-agents and tightly constrained tools. The workflow decomposes software engineering program repair into four components governed by respective sub-agents. This simple, forced decoupling of patch proposing, debugging, patch revision, and optional test generation (1) reduces long-context burden on any individual sub-agent, (2) focuses each sub-agent on specific, pre-defined sub-tasks, and (3) allows for specialized performance improvement on specific sub-tasks. When provided human-written tests, TDFlow attains 88.8% pass rate on SWE-Bench Lite (an absolute improvement of 27.8% over the next best baseline) and 94.3% on SWE-Bench Verified. In this work, we further show that the primary obstacle to human-level software engineering performance lies within writing successful reproduction tests. Manual inspection of the 800 TDFlow runs within SWE-Bench Lite and Verified uncover only 7 instances of test hacking, which were subsequently counted as failures. We envision a human-LLM interactive system powered by TDFlow where human developers write tests solved by LLM systems. Together, these results show that modern LLMs, when embedded in a narrowly engineered, test-driven workflow, already achieve human-level test resolution – with the final frontier for fully autonomous repository repair being accurate reproduction test generation.

Large language models (LLMs) often generate off-domain or harmful responses when deployed in specialized, high-stakes domains, motivating the need for rigorous LLM domain certification. While the VALID algorithm (Emde et al., 2025) achieves formal domain certificate guarantee using a guide model G trained from scratch on in-domain data, it suffers from poor generalization due to limited training. In this work, we propose PRISM, a novel approach that overcomes this key limitation by leveraging pretrained language models as guide models, enhanced via contrastive fine-tuning to sharply distinguish acceptable from refused content. We explore and experiment variants of PRISM with different loss functions to ensure that the model exploits the rich world knowledge of pretrained models while aligned to the target domain. We show that two variants of PRISM, PRISM-BC and PRISM-GA, achieve superior OOD rejection and tighter certification bounds across eight diverse data regimes and perturbations, establishing a more reliable approach to domain-adherent LLM deployment.

We present AutoBool, a reinforcement learning (RL) framework that trains large language models (LLMs) to generate effective Boolean queries for medical systematic reviews. Boolean queries are the primary mechanism for literature retrieval in this domain and must achieve high recall while maintaining reasonable precision—a challenging balance that existing prompt-based LLM approaches often struggle to achieve.A major limitation in this space is the lack of ground-truth best Boolean queries for each topic, which makes supervised fine-tuning impractical. AutoBool addresses this challenge by leveraging RL to directly optimize query generation against retrieval performance metrics, without requiring ideal target queries. To support this effort, we create and release the largest dataset of its kind: 65 588 topics in total for training and evaluating the task of automatic Boolean query formulation.Experiments on our new dataset and two established datasets (CLEF TAR and Seed Collection) show that AutoBool significantly outperforms zero-shot/few-shot prompting and matches or exceeds the effectiveness of much larger GPT-based models (e.g., GPT-4o, O3) using smaller backbones. It also approaches effectiveness of expert-authored queries while retrieving 10–16 times fewer documents. Ablation studies reveal the critical roles of model backbone, size, decoding temperature, and prompt design. Code and data are available at https://github.com/ielab/AutoBool.

Large language models (LLMs) are highly sensitive to subtle changes in prompt phrasing, posing challenges for reliable auditing. Prior methods often apply unconstrained prompt paraphrasing, which risk missing linguistic and demographic factors that shape authentic user interactions. We introduce AUGMENT (Automated User-Grounded Modeling and Evaluation of Natural Language Transformations), a framework for generating controlled paraphrases, grounded in user behaviors. AUGMENT leverages linguistically informed rules and enforces quality through checks on instruction adherence, semantic similarity, and realism, ensuring paraphrases are both reliable and meaningful for auditing. Through case studies on the BBQ and MMLU datasets, we show that controlled paraphrases uncover systematic weaknesses that remain obscured under unconstrained variation. These results highlight the value of the AUGMENT framework for reliable auditing.

This paper investigates the real-world vulnerabilities of audio-based large language models (ALLMs), such as Qwen2-Audio. We first demonstrate that an adversary can craft stealthy audio perturbations to manipulate ALLMs into exhibiting specific targeted behaviors, such as eliciting responses to wake-keywords (e.g., “Hey Qwen”), or triggering harmful behaviors (e.g., “Change my calendar event”). Subsequently, we show that playing adversarial background noise during user interaction with the ALLMs can significantly degrade the response quality. Crucially, our research illustrates the scalability of these attacks to real-world scenarios, impacting other innocent users when these adversarial noises are played through the air. Further, we discuss the transferability of the attack and potential defensive measures.

We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as semantic or stylistic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use.To demonstrate our model’s interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by ~8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes get assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data — enabling generalization even when supervision blends semantic or stylistic factors.

Real-world clinical text-to-SQL requires reasoning over heterogeneous EHR tables, temporal windows, and patient-similarity cohorts to produce executable queries. We introduce ClinSQL, a benchmark of 633 expert-annotated tasks on MIMIC-IV v3.1 that demands multi-table joins, clinically meaningful filters, and executable SQL. Solving ClinSQL entails navigating schema metadata and clinical coding systems, handling long contexts, and composing multi-step queries beyond traditional text-to-SQL. We evaluate 20 proprietary and open-source models under Chain-of-Thought self-refinement and use rubric-based SQL analysis with execution checks that prioritize critical clinical requirements. Despite recent advances, performance remains far from clinical reliability: on the test set, GPT-5-mini attains 74.7% execution score, DeepSeek-R1 leads open-source at 69.2% and Gemini-2.5-Pro drops from 85.5% on Easy to 67.2% on Hard. Progress on ClinSQL marks tangible advances toward clinically reliable text-to-SQL for real-world EHR analytics.

As Large Language Models(LLMs) increasingly power chatbots, social media, and other interactive platforms, the ability to detect AI in conversational settings is critical for ensuring transparency and preventing potential misuse. However, existing detection methods focus on static, document-level content, overlooking the dynamic nature of dialogues. To address this, we propose an utterance-level detection framework, which integrates features from individual and combined analysis of dialogue participants’ responses to detect LLM-generated text under conversational setting. Leveraging a transformer-based recurrent architecture and a curated dataset of human-human, human-LLM, and LLM-LLM dialogues, this framework achieves an accuracy of 98.14% with high inference speed, supported by extensive results of experiments on different models and settings. This work provides an effective solution for detecting LLM-generated text in real-time conversations, promoting transparency, and mitigating risks of misuse.

Reliably ensuring Large Language Models (LLMs) follow complex instructions is a critical challenge, as existing benchmarks often fail to reflect real-world use or isolate compliance from task success. We introduce MOSAIC (MOdular Synthetic Assessment of Instruction Compliance), a modular framework that uses a dynamically generated dataset with up to 20 application-oriented generation constraints to enable a granular and independent analysis of this capability.Our evaluation of five LLMs from different families based on this new benchmark demonstrates that compliance is not a monolithic capability but varies significantly with constraint type, quantity, and position. The analysis reveals model-specific weaknesses, uncovers synergistic and conflicting interactions between instructions, and identifies distinct positional biases such as primacy and recency effects. These granular insights are critical for diagnosing model failures and developing more reliable LLMs for systems that demand strict adherence to complex instructions.

We present STAMP (Selective Task-Aware Mechanism for Text Privacy), a new framework for task-aware text privatization that achieves an improved privacy–utility trade-off. STAMP selectively allocates privacy budgets across tokens by jointly considering (i) each token’s importance to the downstream task (as measured via a task- or query-specific representation), and (ii) its privacy sensitivity (e.g., names, dates, identifiers). This token-level partitioning enables fine-grained, group-wise control over the level of noise applied to different parts of the input, balancing privacy protection with task relevance. To privatize individual token embeddings, we introduce the polar mechanism, which perturbs only the direction of embeddings on the unit sphere while preserving their magnitude. Decoding is performed via cosine nearest-neighbor search, aligning the perturbation geometry with the decoding geometry. Unlike isotropic noise mechanisms, the polar mechanism maintains semantic neighborhoods in the embedding space and better preserves downstream utility. Experimental evaluations on SQuAD, Yelp, and AG News datasets demonstrate that STAMP, when combined with the normalized polar mechanism, consistently achieves superior privacy–utility trade-offs across varying per-token privacy budgets.

We model Semantic Self-Verification (SSV) as the problem of determining whether a statement accurately characterizes its own semantic properties within a given interpretive framework that formalizes a challenge in AI safety and fairness: can an AI system verify that it has correctly interpreted rules intended to govern its behavior? We prove that SSV, in this specification, is NP-complete by constructing a polynomial-time reduction from 3-Satisfiability (3-SAT). Our reduction maps a 3-SAT formula to an instance of SSV involving ambiguous terms with binary interpretations and semantic constraints derived from logical clauses. This establishes that even simplified forms of semantic self-verification should face computational barriers. The NP-complete lower bound has implications for AI safety and fairness approaches that rely on semantic interpretation of instructions, including but not limited to constitutional AI, alignment via natural language, and instruction-following systems. Any approach where an AI system verify its understanding of directives faces this computational barrier. We argue that more realistic verification scenarios likely face even greater complexity.