TACL 2026 Papers

While large language models (LLMs) are widely used across cultures, they often generate culturally inappropriate responses in unfamiliar cultural contexts due to biases embedded in their training data. Existing approaches primarily rely on expanding static cultural knowledge, which fails to capture the inherently relative and context-dependent nature of culture. In this paper, we propose a Cultural value-based Reasoning (CURE) framework that interprets behaviors through underlying cultural value systems. In addition, we integrate CURE into LLMs via Chain-of-Thought (CoT) distillation, referred to as CURE-distillation, to internalize culturally grounded reasoning. Experimental results show that models trained with CURE-distillation improve cultural adaptability, enabling them to produce culturally aligned ethical judgments across diverse cultural scenarios. These results suggest that strengthening sociocultural reasoning capabilities can substantially improve the cultural adaptability of LLMs. The code is available at https://github.com/KUNLP/CURE.

Conspiracy theories erode public trust in science and institutions while resisting debunking by evolving and absorbing counter-evidence. As AI-generated misinformation becomes increasingly sophisticated, understanding the rhetorical patterns in conspiratorial content is important for developing interventions such as targeted prebunking and assessing AI vulnerabilities. We introduce CONSPIRED (CONSPIR Evaluation Dataset), which captures the cognitive traits of conspiratorial ideation in multi-sentence excerpts (80–120 words) from online conspiracy articles, annotated using the CONSPIR cognitive framework. CONSPIRED is the first dataset of conspiratorial content annotated for general cognitive traits. Using CONSPIRED, we (i) develop computational models that identify conspiratorial traits and the dominant trait in text excerpts, and (ii) evaluate LLM robustness to conspiratorial inputs. We find that LLMs are readily misaligned by conspiratorial framing, reproducing its rhetorical patterns even when successfully deflecting comparable fact-checked misinformation.1

Despite recent breakthroughs in reasoning-enhanced large language models (LLMs), incorporating inference-time reasoning into application tasks such as machine translation (MT), where human translators naturally employ structured, multi-layered reasoning chain-of-thoughts (CoTs), is yet un-derexplored. Existing methods either design a fixed CoT tailored for a specific MT sub-task (e.g., literature translation), or rely on synthesizing CoTs unaligned with humans and supervised fine-tuning (SFT) prone to overfitting, limiting their adaptability to diverse translation scenarios. This paper introduces R1-Translator (R1-T1), a novel framework to achieve inference-time reasoning for general MT via reinforcement learning (RL) with human-aligned CoTs comprising six common patterns. Our approach pioneers three innovations: (1) verifying reasoning-based translation in various MT scenarios (e.g., multilingual MT, domain MT) unseen from the training phase; (2) formalizing six expert-curated CoT templates that mirror hybrid human strategies like context-aware paraphrasing and round-trip translation; and (3) enabling more flexible CoTs through an RL stage after cold-start. Both human and automatic evaluation results indicate a steady translation quality improvement in a total of 10+ languages and 40+ translation directions on Flores-101 test set and four domain-specific MT tasks, especially on the languages unseen from training.

Code-capable large language model (LLM) agents are embedded in software engineering workflows where they can read, write, and execute code, raising “jailbreak" stakes beyond text-only settings. Prior evaluations emphasize refusal or harmful-text detection, leaving open whether agents compile and run malicious programs. We present JAWS-BENCH(Jailbreaks Across WorkSpaces), a benchmark spanning three escalating workspace regimes mirroring attacker capability: empty (JAWS-0), single-file (JAWS-1), and multi-file (JAWS-M). We pair it with a hierarchical, executable-aware Judge Framework that tests (i) compliance, (ii) attack success, (iii) syntactic correctness, and (iv) runtime executability to measure de-ployable harm. Across seven LLM backends from five families, prompt-only attacks in JAWS-0 achieve 61% compliance; 58% are harmful, 52% parse, and 27% run end-to-end. In JAWS-1, compliance reaches 100% for stronger models with a mean ASR (Attack Success Rate) ≈ 71%; JAWS-M raises mean ASR to ≈ 75%, with 32% runnable attack code. Wrapping an LLM in an agent increases ASR by 1.6×, by overturning initial refusals during planning and tool use. Additional evaluations with SWE-Agent and OpenAI Codex exhibit similar trends, indicating that JAWS-BENCH can be reused across multiple agent frameworks. Category analyses identify which attack classes are most vulnerable and deployable, motivating execution-aware defenses and refusal-preserving agent designs.

Human reasoning often involves working over limited information to arrive at probabilistic conclusions. In its simplest form, this involves making an inference that is not strictly entailed by a premise, but rather only likely given the premise. While reasoning LLMs have demonstrated strong performance on logical and mathematical tasks, their behavior on such open-ended, non-deterministic inferences remains largely unexplored. We introduce PROBCOPA, a dataset of 210 handcrafted probabilistic inferences in English, each annotated for inference likelihood by 25–30 human participants. We find that human responses are graded and varied, revealing probabilistic judgments of the inferences in our dataset. Comparing these judgments with responses from eight state-of-the-art reasoning LLMs, we show that models consistently fail to produce human-like distributions. Finally, analyzing LLM reasoning chains, we find evidence of a common reasoning pattern used to evaluate such inferences. Our findings reveal persistent differences between humans and LLMs, and underscore the need to evaluate reasoning beyond deterministic settings.1

As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model’s post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model’s values, and subsequent preference optimization rarely realigns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.1

State-of-the-art NLP benchmarks require interpretation of natural-language that specify conditions, procedures, and exceptions, often relying on implicit assumptions and external knowledge. Constructing complete semantic representations with proof-theoretic guarantees is frequently impractical at scale, and purely text-based reasoning offers limited means of inspection. This paper asks how much understanding of benchmark language can be achieved when formal semantic guarantees are weakened. We investigate this question by extracting computables: executable representations whose runtime behavior provides operational evidence of semantic adequacy, including executability, execution traces, and runtime failures. We induce and iteratively refine computables for benchmark instances using retrieval from external knowledge. Across mathematical reasoning, multi-step reasoning, causal inference, and rule- and exception-heavy legal and biomedical benchmarks, we find that the proposed approach consistently exceeds text-only reasoning and one-shot code execution. Beyond accuracy, our analyses show that these computables provide scalable, inspectable semantic evidence: they expose conditions and exceptions benchmark language forces into executable form, offering a practical bridge between proof-oriented semantics and purely textual reasoning.

Numerical reasoning over expert-domain tables often exhibits high in-domain accuracy but limited robustness to domain shift. Models trained with supervised fine-tuning (SFT) on specific datasets tend to rely on header–operation shortcuts rather than structural reasoning. We introduce TaNOS, a continual pretraining framework comprising three components: (i) header anonymization to reduce lexical memorization, (ii) operation sketches that provide minimal structural cues, and (iii) self-supervised pretraining that constructs correctness-guaranteed program–question pairs from given tables in a program-first manner. By decoupling domain semantics and numerical operation structure, TaNOS improves the transferability of numerical reasoning. Applied to an 8B instruction-tuned model, TaNOS achieves 80.13% execution accuracy on FinQA with only 10% train data, outperforming SFT baseline (73.97%) with full train data and proprietary models such as GPT-5, Gemini-2.5-Pro. Furthermore, in the domain-shift experiments, TaNOS displays nearly-negligible cross-domain gap (<2pp) when standard SFT shows over 10pp gap. These results suggest that structural guidance with operation sketches, header-agnostic representations, and correctness-guaranteed self-supervision can improve the robustness of numerical reasoning across diverse expert-domain tables.

As Large Language Models (LLMs) expand beyond text, integrating speech as a native modality has given rise to SpeechLLMs, which directly process spoken language and enable speech-to-text translation (ST) and other downstream tasks, bypassing traditional transcription-based pipelines. Whether this integration improves ST quality over established cascaded architectures, however, remains an open question. We present Hearing to Translate,1 the first comprehensive test suite rigorously benchmarking 6 state-of-the-art SpeechLLMs against 16 strong direct and cascade systems that couple leading speech foundation models (SFM), with multilingual LLMs. Our analysis spans 16 benchmarks, 13 language pairs, and 9 challenging conditions, including disfluent, noisy, and long-form speech. Across this extensive evaluation, we find that cascaded systems remain the most reliable solution overall, but most recent SpeechLLMs can match or even outperform cascades in various settings while SFMs lag behind both, highlighting that integrating an LLM, either within the model or in a pipeline, is essential for high-quality speech translation.

Large Language Models (LLMs) are effective at deceiving when prompted to do so. Models that demonstrate better performance on reasoning tasks are also better at prompted deception. But under what conditions do they deceive without instruction to do so? This study evaluates unsolicited deception produced by LLMs in a preregistered experimental protocol using tools from signaling theory. We evaluated a range of 18 proprietary closed-source and open-source LLMs using modified 2x2 games (in the style of the Prisoner’s Dilemma) augmented with a phase in which they can freely communicate to the other agent using unconstrained language. This setup creates an opportunity to misrepresent its actions in conditions that vary in how useful doing so might be towards goal satisfaction. The results indicate that 1) all tested LLMs misrepresent their actions in at least some conditions, 2) they are generally more likely to do so in situations in which deception is beneficial, and 3) models exhibiting better reasoning capacity overall tend to misrepresent at higher rates. Taken together, these results suggest a correlational relationship between model reasoning performance and situational deception, and reveal certain contextual factors that affect whether LLMs will misrepresent actions or not in a novel experimental configuration.

Statistical analysis of corpora provides an approach to quantitatively investigate natural languages. This approach has revealed that several power laws consistently emerge across different corpora and languages, suggesting universal mechanisms underlying languages. In particular, the power-law decay of correlations has been interpreted as evidence of underlying hierarchical structures in syntax, semantics, and discourse. This perspective has also been extended beyond corpora produced by human adults, including child speech, birdsong, and chimpanzee action sequences. However, the argument supporting this interpretation has not been empirically tested in natural languages. To address this gap, the present study examines the validity of the argument for syntactic structures. Specifically, we test whether the statistical properties of parse trees align with the assumptions in the argument. Using English and Japanese corpora, we analyze the mutual information, deviations from probabilistic context-free grammars (PCFGs), and other properties in natural language parse trees, as well as in the PCFG that approximates these parse trees. Our results indicate that the assumptions do not hold for syntactic structures and that it is difficult to apply the proposed argument not only to sentences by human adults but also to other domains, highlighting the need to reconsider the relationship between the power law and hierarchical structures.

Large reasoning models (LRMs) have shown exceptional performance in complex tasks such as mathematics and coding. In the field of machine translation (MT), reinforcement learning (RL) has been utilized to enhance the quality of translations. However, traditional RL approaches rely heavily on the base model’s inherent translation capabilities, which may falter when dealing with terminology translations and domain-specific expressions without sufficient guidance. In this paper, we introduce FeedTrans (Feedback-driven Translation), which employs a FeedRollout mechanism to incorporate feedback as guidance, enabling the production of high-quality translations and expanding the search space for improved translation outcomes. Extensive experiments across six translation tasks validate the effectiveness of our approach. To the best of our knowledge, this is the first attempt at utilizing feedback in MT-oriented RL. https://github.com/Acerkoo/FeedTrans

Test-Time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, In-Context Learning has been widely explored at inference time to enhance model performance without weight updates. However, TTRL’s two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose Context-Guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7% relative accuracy improvement over TTRL), while boosting efficiency by obtaining strong performance after only a few steps of Test-Time Training (e.g. 8% relative improvement rather than 1% over TTRL after 3 steps).

Group Relative Policy Optimization (GRPO) has been shown to be an effective algorithm when an accurate reward model is available. However, such a highly reliable reward model is not available in many real-world tasks. In this paper, we particularly focus on multi-objective settings, in which we identify that GRPO is vulnerable to reward hacking, optimizing only one of the objectives at the cost of the others. To address this issue, we propose MO-GRPO, an extension of GRPO with a simple normalization method to reweight the reward functions automatically according to the variances of their values. We first show analytically that MO-GRPO ensures that all reward functions contribute evenly to the loss function while preserving the order of preferences, eliminating the need for manual tuning of the reward functions’ scales. Then, we evaluate MO-GRPO experimentally in three domains: (i) the multi-armed bandits problem, (ii) machine translation tasks on the WMT benchmark (En-Ja, En-Zh), and (iii) the instruction following task. MO-GRPO achieves stable learning by evenly distributing correlations among the components of rewards, outperforming GRPO, showing MO-GRPO to be a promising algorithm for multi-objective reinforcement learning problems.

Figurative language and negation are two areas that challenge current language models, however, both are widely used throughout written and spoken language. Large language models (LLMs) are also widely used in everyday contexts where they cannot necessarily be tuned for a specific dataset. It is therefore essential to understand the ability of LLMs to correctly interpret text that includes both negation and figurative language. To investigate this, we develop a set of new annotations to an existing dataset of figurative language, and test a range of language models on the dataset. We find that the combination of negation and figurativeness can present a particular challenge, and that performance overall and across different negation types is particularly dependent on the prompt style used.

Humans have an uncanny ability to push words beyond their limits. This ability manifests in many phenomena, including metaphor, metonymy, semantic drift, slang, jargon, conversion, and overextension. Generative models of these phenomena are uncommon, because there is no robust methodology that can be used to train and evaluate models of this nature. To address this, we introduce a new task, novel sense formulation, in which a model is exposed to a multimodal representation of an unseen concept and must use an existing word creatively to describe it. We create seven datasets corresponding to the phenomena above, and evaluate a perceptron, a transformer, and an influential cognitive model. For most phenomena, a multimodal variant of the perceptron performed best. The cognitive model underperformed, suggesting that its description of the mechanism behind sense extension is incomplete.

As the context window of Large Language Models (LLMs) continues to expand, the data required to effectively train and evaluate these capabilities remains underexplored. With existing research primarily focuses on architectural optimization, there is a need for a systematic, data-centric review. This survey bridges this gap by investigating the data foundations of Long-Context Language Models (LCMs). We begin by examining current data strategies alongside their strengths and limitations, mapping the required data to desired model capabilities. Building on this, we explore how targeted training data designs drive core, often interconnected skills such as retrieval, reasoning, and aggregation. Furthermore, we analyze the evaluation landscape, illustrating how selecting appropriate benchmarks is crucial for probing capability boundaries and guiding effective model selection. Finally, we synthesize actionable guidelines for data construction and outline critical future directions to propel the advancement of long-context language models, including quantifying data quality, establishing scaling laws for length distributions, and developing dynamic evaluation frameworks.

Tokenization remains a persistent bottleneck in language modeling, especially when vocabulary learning is limited by whitespace boundaries. We present SupraTok, a tokenizer that crosses whitespace boundaries using three modular components: optional entropy-based data curation, staged curriculum training with PMI-guided candidate search, and multilingual script handling. At 100k vocabulary on the same unfiltered training data, SupraTok improves compression over standard BPE by 17.5% and over the official SuperBPE implementation by 1.8%, while training 2.1× faster than SuperBPE. Across 50k–300k vocabularies in the same matched setting, Supra-Tok remains ahead of SuperBPE by 1.8%–8.6%. We evaluate entropy filtering separately as a pipeline step: at 100k vocabulary it raises SupraTok from 5.78 to 5.99 C/T, while matched controls show a smaller gain for SuperBPE and almost no change for SP-BPE-CrossBoundary. On FLORES-200 across 14 languages, SupraTok yields a macro-averaged 34.9% relative gain over the BPE baseline. In separate downstream experiments with matched compute and fixed token budgets, using 12L-768d and 24L-1024d GPT-2-style backbones with 256k vocabularies, SupraTok improves HellaSwag and MMLU. Overall, these results show that crossing whitespace boundaries gives consistent compression gains under controlled public-data comparisons, while optional entropy filtering provides a separate pipeline benefit.

Large Language Models (LLMs) exhibit unique strengths arising from differences in model architecture, training data, and strategies. Ensemble learning has been explored to leverage these complementary strengths through decision-level sharing (i.e.,Decision Ensemble), which combines the predictions from multiple LLMs. However, such methods integrate only shallow decisions and overlook the exchange of deeper levels of information within the internal representations of LLMs, such as problem understanding, world knowledge, and latent reasoning patterns. In this work, we propose Representation Ensemble (RISE), a novel ensemble framework that enables cross-LLM representation sharing for richer information exchange. To address challenges of representation-level interaction caused by layer misalignment and latent-space incompatibility across LLMs, we introduce a representation alignment method based on relational similarity measures and an orthogonal latent-space transformation. Experimental results show that (1) RISE achieves performance competitive with existing decision ensemble methods, and (2) RISE is strongly complementary to decision ensemble, with their combination boosting collaboration gains by 14%–41%. Finally, we further compare ensemble of small LLMs to a single larger LLM and to model merging and composition approaches, and find that ensemble learning consistently generalizes well without additional training.