United States

Thinking Mode Fusion (TMF) enables large language models to support both concise responses and long-form reasoning by unifying a non-thinking mode and a thinking mode within a single model. However, its training dynamics, including the \emph{data ratio} and \emph{training schedule} between the two modes, remain underexplored. In this work, we present a systematic study of TMF by analyzing the effects of the training schedule and data ratio between thinking and non-thinking modes. Focusing on mathematical problem solving, we construct a benchmark with multiple thinking-to-non-thinking data ratios and three training schedules. Our results reveal an asymmetric interaction between the two modes: increasing the ratio of non-thinking supervision reduces the accuracy of the thinking mode. We further show that different training schedules modulate this trade-off and that the optimal schedule depends on the data ratio. Finally, we quantify a negative correlation between non-thinking and thinking mode supervision, highlighting an inherent tension between these two modes. These findings provide practical guidance for designing effective TMF training settings. All code and data are released to support further research at: \href{https://anonymous.4open.science/r/Fusion-Bench-2D62}{\textbf{Fusion Bench}}.

ACL 2026

Fusion Training for Mathematical Generalization in Large Language Models

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

poster

We ask whether topic sentiment has a causal effect on perceived political ideology, and whether the answer depends on who assigns the ideology label. Using the AllSidesV2 test dataset (N=11,516 articles), we hold sentiment annotations constant and compare ideology labels from expert human annotators, GPT-4o-mini (baseline and finetuned), and Llama-3.3-70B. We apply Double Machine Learning (DML) and community-level mediation analysis across all four annotation paradigms. Human annotations yield no significant causal effects at the community level. Fine-tuned GPT-4o-mini achieves the highest classification accuracy (F1=72.48) and is the only annotator paradigm that produces significant community-level treatment effects and significant natural direct effects (NDEs) in mediation. We interpret this as evidence of shortcut learning: fine-tuning on ideology-labeled data causes the model to internalise a spurious sentiment--ideology coupling not operative in human judgment. This coupling is structurally invisible to F1-based evaluation, with implications for the use of LLM annotations as silver labels and as proxies for human judgment in downstream causal analyses.

Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles

In conversational implicatures, speakers convey hidden intended meanings beyond the literal content of their utterances, and hearers are expected to infer what is implied. This study examines how Large Language Models (LLMs) interpret conversational implicatures, using human interpretation as a baseline and gold standard for comparison. The same experiments were conducted with two types of participants: humans and LLMs. Two metrics were adopted: a surprisal-based metric and a response-based metric. The results suggest that the response-based metric demonstrates higher accuracy, comparable to human responses, than the surprisal-based metric. In particular, humans and LLMs using the response-based metric performed better in the literal condition than in the implied condition. Additionally, they were more sensitive to capturing implied meanings for some-all trigger than for the other triggers. Overall, LLMs employing the response-based metric tend to exhibit human-like behavior, but still diverge from humans in their understanding of conversational implicatures.

Understanding Conversational Implicatures in Humans and LLMs

Masked language modeling for low-resource ancient languages remains challenging because pre-trained multilingual models lack exposure to these languages. We investigate rule-based linguistic constraints and hard negative mining for Sumerian, a language isolate absent from multilingual BERT's training data. We build a hierarchical validator capturing subword, word, and part-of-speech patterns from 4,545 annotated sequences, using it to filter candidates and identify hard negatives for fine-tuning. Vanilla mBERT achieves 18.0\% hit@10 accuracy. The validator alone improves this to 72.8\%, while hard negative fine-tuning reaches 78.3\%. Combining both yields 86.7\%, a 68.7 percentage point improvement. Temporal generalization evaluation on tablets from 600 years earlier shows that both the hard negative mining and the validator alone improve performance, but the combined approach underperforms due to the validator's period specific rules. These findings demonstrate that hard negative mining transfers across periods while explicit rule-based constraints provide strong in-domain improvements but limited cross-period generalization.

Validator-Guided Hard Negative Mining for Masked Language Modeling in Low-Resource Ancient Languages

LLM routing directs queries to a cheaper model when it suffices and to an expensive model otherwise, reducing inference cost. Existing input-based routers optimize cost-performance trade-offs but provide no formal bound on how often the cheaper model fails among routed queries. We adapt a proactive conformal gate framework to LLM routing. A logistic regression gate trained on text embeddings predicts per-query safety, and Clopper-Pearson conformal calibration selects a routing threshold that guarantees the violation rate among routed queries stays below $\alpha$ (the violation tolerance) with probability at least $1 - \delta$ (the confidence level). On two benchmarks covering math reasoning (GSM8K) and multi-domain knowledge (MMLU), routing between Mixtral-8x7B and GPT-4 (a 24.5$\times$ cost difference), our method maintains the target $\alpha$ within the $\delta$ tolerance across a sweep from 0.05 to 0.50, while a validation-tuned baseline crosses the violation boundary on GSM8K. A feasibility analysis across all 10 RouterBench models reveals that routability is jointly model- and task-dependent. To our knowledge, this is the first input-based LLM router with distribution-free safety guarantees.

Conformal LLM Routing with Distribution-Free Safety Guarantees

Turkish idiomatic light verb constructions (LVCs) are challenging for multiword expression processing because they often share the same surface form as fully literal verb–object combinations while functioning as a single, partially idiomatic predicate. We frame Turkish LVC detection as a binary classification task (literal ditransitive vs. LVC) and evaluate on a manually created controlled set (N=147) with matched negatives: out-of-domain random sentences and in-domain literal controls (NLVC), alongside LVC positives. We compare a supervised Turkish encoder baseline (BERTurk with a classifier head) to three instruction-tuned LLMs from different families under zero-shot, one-shot, and few-shot prompting, and analyze how demonstrations shift error profiles. In zero-shot, LLMs perform well on negatives but show very low LVC recall. One-shot prompting sharply improves LVC detection but can induce strong, model-specific biases (over- vs. under-predicting LVC). A richer master few-shot prompt improves calibration and yields robust overall performance for GPT-OSS-20B and Qwen 2.5. Overall, the results highlight substantial prompt sensitivity in Turkish metalinguistic classification: the supervised baseline remains competitive, while prompted LLMs can match or exceed it on LVCs with carefully constructed demonstrations. We will release code, prompts, and evaluation materials to support reproducibility.

Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification

Large language models (LLMs) are increasingly deployed in high-stakes settings, yet reliably estimating when their outputs should be trusted remains an open challenge. Existing uncertainty estimation approaches---such as calibration, token-level probabilities, or semantic entropy---typically require access to model internals, additional supervision, or computationally intensive pipelines. We propose answer instability, defined as the variability of a model's final answer across repeated stochastic generations of the same prompt, as a simple, label-free, and black-box uncertainty signal. Evaluated across three task types — reasoning, multiple-choice QA, and constraint-following — using four LLMs and 520 prompt-model pairs, our approach achieves performance competitive with semantic entropy while requiring no semantic similarity model. Our results show that instability strongly correlates with prediction errors and reliably discriminates correct from incorrect outputs. We further demonstrate its utility for selective prediction and targeted repair, improving reliability without access to internal probabilities or additional training.

When Models Hesitate: Answer Instability as a Label-Free Uncertainty Signal for LLMs

Learning reward functions from preference feedback is a widely used strategy for aligning intelligent systems with human intent, but its effectiveness heavily relies on the representation used to describe candidate outcomes. In perceptual domains such as visual aesthetics and robot appearance, hand-engineered features often fail to capture the latent attributes underlying subjective judgments. We propose **HARP**, a representation-based preference learning framework that infers reward functions directly over learned latent embeddings using pairwise comparisons and Bayesian updates, interpreting preference learning as identifying a preference-aligned direction in representation space. By operating in a continuous representation space, HARP enables preference learning without manual feature design. We evaluate HARP on a robot image preference task with simulated oracle users, comparing learned representations to principled hand-engineered features under identical query strategies. Results show more stable convergence and improved top-$k$ preference recovery, highlighting the importance of representation choice in perceptual preference learning.

HARP: Representation-Based Preference Learning for Perceptual Data

Humans play a vital role at every stage of AI development, from data collection and curation to model development and evaluation. However, humans often disagree with each other and sometimes with themselves over time. It is essential to take disagreement into account when building human-centered AI systems, especially in domains where it is prevalent, such as AI safety, content moderation, or sentiment analysis. Disagreement typically arises from subjective human opinion and can vary with one's identity, beliefs, and social environment. Despite this, current LLM evaluation approaches frequently rely on aggregating labels (often via plurality voting) to represent consensus, failing to account for variance in human responses. This amplifies the reproducibility crisis in AI evaluation. Human feedback is also crucial for ensuring that AI systems align with human values. For these systems to be trustworthy, it is critical to ensure that they reflect diverse human values and perspectives. In this thesis proposal, we present a human-centered and perspective-aware framework for reproducible ML evaluation and AI alignment.

Thesis Proposal: Toward a Human-Centered and Perspective-Aware Framework for Reproducible ML Evaluation and AI Alignment

In-person cognitive assessments are structured clinical conversations that provide important cues about patients' cognition but are difficult to analyze at scale. Existing datasets do not capture both speaker roles and Dialogue-Acts in this setting, limiting progress in modeling conversational structure. We introduce a dataset of anonymized cognitive assessment transcripts annotated with speaker roles and Dialogue-Act labels. We evaluate large language models on two tasks: dialogue-act classification and patient utterance generation. Our results show that while few-shot prompting improves performance, models still struggle to capture fine-grained pragmatic intent and produce clinically consistent responses. These findings highlight a gap between fluent text generation and true dialogue understanding in clinical settings.

Understanding Clinical Cognitive Dialogues Using Large Language Models

Deploying machine learning models in real-world domain-specific scenarios is challenged by the scarcity of expert annotations and by data drift, where the statistical properties of incoming data continuously evolve. Active Learning (AL) iteratively improves compact models with expert annotations but suffers from recurring cold-start degradation, while LLMs provide strong off-the-shelf performance yet cannot leverage newly accumulated labels, raising the question: how can we better leverage LLMs to assist the active learning process? Through an empirical study on five legal and biomedical datasets, we reveal a complementary temporal dynamic: LLMs excel during early and post-drift stages, while AL-assisted compact models eventually surpass them as annotations accumulate. Motivated by this finding, we propose an ensemble system that combines an LLM, an AL-assisted compact model, and an automatic switch module that routes predictions to the better-performing model in real time. Evaluated under simulated data drift on two mental health datasets, our system achieves 96--98% switch accuracy and consistently outperforms either model used alone.

Downloads

Next from ACL 2026

Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles

Similar lecture

Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles

Similar lecture

Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning

Downloads