United States

Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resource languages. We attribute this issue as a mismatch gap between language-agnostic semantic understanding ability and language dominant safety alignment biased toward high-resource languages. Based on above insights, we empirically identify the semantic bottleneck in LLMs: intermediate layers in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Then, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwen3 Instruct models (7B–32B). Besides, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model&#39;s language-agnostic semantic space.

ACL 2026

LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<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://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<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 EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

EACL 2026

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resource languages. We attribute this issue as a mismatch gap between language-agnostic semantic understanding ability and language dominant safety alignment biased toward high-resource languages. Based on above insights, we empirically identify the semantic bottleneck in LLMs: intermediate layers in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Then, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwen3 Instruct models (7B–32B). Besides, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model's language-agnostic semantic space.

poster

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.

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

Multilingual large language models (LLMs) are expensive to pretrain and often suffer from imbalances across languages and datasets, English-centric bias, tokenizer oversegmentation for morphologically rich low-resource languages, and the curse of multilinguality. We introduce PARAMANU, a family of Indian language-only autoregressive language models trained from scratch on open-source language-specific data for the five most spoken Indian languages: Bangla (Bengali), Hindi, Marathi, Tamil, and Telugu. All models are designed for affordability and are trained on a single GPU with a budget under $1,000, allowing under-resourced researchers to build competitive language models. To address low-resource challenges, we develop morphology-aligned, low-fertility tokenizers, and propose an interpolation-based method for token position indices in RoPE based scaling to train longer sequences efficiently. We also create instruction-tuning datasets in Bangla that are then translated to the other four languages. Despite their small size (108M-367M parameters), Paramanu achieves a strong performance-efficiency tradeoff and outperforms most larger multilingual models up to 8B across all five languages. 
The models and datasets are available at: https://huggingface.co/collections/mitodru/paramanu.

Paramanu: Compact and Competitive Monolingual Language Models for Low-Resource Morphologically Rich Indian Languages

KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls that practitioners should be aware of when deploying KV cache compressed LLMs. We evaluate five KV cache compression methods (StreamingLLM, SnapKV, TOVA, H2O, and K-Norm) on Llama3.1 8B and Qwen2.5 14B under multi-instruction prompting with IFEval. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example, we highlight system prompt leakage as a case study, empirically demonstrating the impact of compression on leakage and general instruction-following. We identify several factors that contribute to system prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.

The Pitfalls of KV Cache Compression

Training language models and examining their linguistic behaviors have been a common protocol in computational linguistics for studying linguistic phenomena and modeling human language processing. However, work in this area is often limited to proof-of-concept demonstrations with arbitrary model configurations, without considering hyperparameter sensitivity, an important source of variation in model performance. In this work, we replicate three prior studies (Chang and Bergen, 2022; Hu et al., 2020b; Kuribayashi et al., 2024) with hyperparameters varied within a practical range, and show that modest hyperparameter changes can alter some qualitative conclusions about models' linguistic abilities and even reverse the ranking of model performance. Our results highlight the risk that prior work may have reflected optimization artifacts rather than the genuine inductive biases of model classes, and that hyperparameter sensitivity should receive more attention as a factor that can meaningfully influence model behavior. We suggest future work to report the variation of performance across the configuration space to enhance the reliability and generalizability of conclusions. 

Code: https://github.com/compling-wat/tune-linguistic-lms.

On the Effect of Hyperparameters in Language Modeling for Computational Linguistics

Production LLM systems often rely on separate models for safety and other classification-heavy steps, increasing latency, VRAM footprint, and operational complexity. We instead reuse computation already paid for by the serving LLM: we train lightweight probes on its hidden states and predict labels in the same forward pass used for generation. We frame classification as representation selection over the full token$\times$layer hidden-state tensor, rather than committing to a fixed token or fixed layer (e.g., first-token logits or final-layer pooling). To implement this, we introduce a two-stage aggregator that (i) summarizes tokens within each layer and (ii) aggregates across layer summaries to form a single representation for classification. We instantiate this template with direct pooling, a 100K-parameter scoring-attention gate, and a downcast multi-head self-attention (MHA) probe with up to 35M trainable parameters. Across safety and sentiment benchmarks our probes improve over logit-only reuse (e.g., MULI) and are competitive with substantially larger task-specific baselines, while preserving near-serving latency and avoiding the VRAM and latency costs of a separate guard-model pipeline. Multi-backbone experiments on dense and mixture-of-experts architectures (Llama-3.2-3B, GPT-OSS-20B, Qwen3-30B-A3B) confirm that these findings generalize beyond a single model family.

A BERTology View of LLM Orchestrations: Token- and Layer-Selective Probes for Efficient Single-Pass Classification

Large Language Model agents can seemingly
plan and act, yet their language use is often
treated primarily as an interface for instructing
actions and reporting results. We argue that this
framing is one important cause of recurrent coordination failures in human-facing and multiagent settings, including ungrounded assumptions, silent goal misalignment, brittle protocol
adherence, and failures to maintain or update
shared dialogue state over time, a limitation previously linked to the absence of explicit common ground tracking in collaborative systems
(Geib et al., 2022). Drawing from classical dialogue system research on joint action, common
ground, grounding, repair, and incremental processing, we re-frame dialogue as part of the
planning loop itself (rather than its output). We
distill this re-framing into concrete implications
for agentic architecture and evaluation, including explicit representations of shared commitments, clarification as a first class action available to the policy, and process metrics that approximate grounding behavior, repair, and commitment formation rather than task completion
alone. We lastly discuss how dialogue-centered
requirements can inform standards and governance for safe deployment of agentic systems.

Dialogue is the Plan: From Interface to Joint Action in Agentic AI

Post-training compression of Multimodal LLMs faces a fundamental geometric conflict: parameter subspaces optimized for text often suppress orthogonal visual features. We demonstrate that standard SVD fails to resolve this cross-modal mismatch, causing catastrophic visual degradation. To bridge this gap, we introduce Joint-Whitening SVD (JW-SVD), a dual-objective framework that aligns vision and language manifolds via a Joint Covariance basis, preserving features critical to both. Additionally, we propose Global Spectrum-Aware Truncation to dynamically transfer parameter budget from the redundant Vision Tower to the sensitive Backbone. Experiments on Qwen2.5-VL and Llama-3-Next confirm that JW-SVD demonstrates superior retention of both text and image capabilities. In addition, it resolves the modality trade-off: it recovers over 30% of perceptual performance lost by baselines while maintaining parity in textual reasoning, enabling robust multimodal performance even at extreme compression rates.

JW-SVD: Bridging the Cross-Modal Mismatch in Post-Training MLLM Compression

This paper presents AutoRAN, the first framework to automate the hijacking of internal safety reasoning in large reasoning models (LRMs). At its core, AutoRAN pioneers an execution simulation paradigm that leverages a weaker but less-aligned model to simulate execution reasoning for initial hijacking attempts and iteratively refine attacks by exploiting reasoning patterns leaked through the target LRM's refusals. This approach steers the target model to bypass its own safety guardrails and elaborate on harmful instructions. We evaluate AutoRAN against state-of-the-art LRMs, including GPT-o3/o4-mini and Gemini-2.5-Flash, across multiple benchmarks (AdvBench, HarmBench, and StrongReject). Results show that AutoRAN achieves approaching 100% success rate within one or few turns, effectively neutralizing reasoning-based defenses even when evaluated by robustly aligned external models. This work reveals that the transparency of the reasoning process itself creates a critical and exploitable attack surface, highlighting the urgent need for new defenses that protect models' reasoning traces rather than merely their final outputs.

AutoRAN: Automated Hijacking of Safety Reasoning in Large Reasoning Models

Reinforcement Learning (RL) has become essential for eliciting complex reasoning capabilities in Large Language Models (LLMs). However, the substantial memory overhead of storing Key-Value (KV) caches during long-horizon rollouts acts as a critical bottleneck, often prohibiting efficient training on limited hardware. While existing KV compression techniques offer a remedy for inference, directly applying them to RL training induces a severe policy mismatch, leading to catastrophic performance collapse. To address this, we introduce Sparse-RL, which empowers stable RL training under sparse rollouts. We show that instability arises from a fundamental policy mismatch among the dense old policy, the sparse sampler policy, and the learner policy. To mitigate this issue, Sparse-RL incorporates Sparsity-Aware Rejection Sampling and Importance-based Reweighting to correct the off-policy bias introduced by compression-induced information loss. Experimental results show that Sparse-RL reduces rollout overhead compared to dense baselines while preserving the performance. Furthermore, Sparse-RL inherently implements sparsity-aware training, significantly enhancing model robustness during sparse inference deployment.

 Sparse-RL: Breaking the Memory Wall in LLM Reinforcement Learning via Stable Sparse Rollouts

Multi-party social dialogue remains underexplored in the literature,
in part due to the difficulty and cost of evaluation. As a result,
recent work on synthetic dialogue generation often relies on automated
metrics and LLM-as-a-Judge frameworks, despite limited evidence that
such judges reflect human preferences in social settings. In this work,
we introduce a lightweight and controllable multi-party dialogue
generation framework (\textsc{MPOD}) as an experimental instrument for
studying generation and evaluation in social interaction. Using this
framework, we conduct human evaluations of open-domain multi-party
dialogue simulation and directly compare human judgments against
state-of-the-art LLM judges. Across 319 pairwise comparisons, we
observe near-random agreement between humans and automated judges
(Cohen's $\kappa \approx 0.11$), driven by systematic behaviors
including extreme tie aversion and strong sensitivity to
assistant-style verbosity. Crucially, human–human inter-annotator
agreement ($\kappa = 0.29$) is substantially higher than human–LLM
agreement. To isolate the
mechanism underlying this misalignment, we introduce a controlled
\textit{Transplant Ablation}, showing that LLM judges consistently
prefer conversations containing a single proprietary, assistant-style
agent. Additional stress tests show that judges prefer GPT-style
conversations even when utterance order is randomly shuffled,
indicating insensitivity to conversational structure and coherence.
Our findings provide controlled evidence that current
instruction-tuned LLM judges do not reliably reflect human preferences for naturalness, engagingness, and overall quality in multi-party social dialogue, calling into question their widespread
use for validating synthetic conversational data.

 Simple Agents, Biased Judges: Efficient Multi-Party Dialogue Generation & The Evaluation Gap

Despite the growing use of large language models (LLMs) for writing tasks, users may hesitate to rely on LLMs when personal style is important. Post-editing LLM-generated drafts or translations is a common collaborative writing strategy, but it remains unclear whether users can effectively reshape LLM-generated text to reflect their personal style. We conduct a pre-registered online study ($n=81$) in which participants post-edit LLM-generated drafts for writing tasks where personal style matters to them. Using embedding-based style similarity metrics, we find that post-editing increases stylistic similarity to participants' unassisted writing and reduces similarity to fully LLM-generated output. However, post-edited text still remains stylistically closer in style to LLM text than to participants' unassisted control text, and it exhibits reduced stylistic diversity compared to unassisted human text. We find a gap between perceived stylistic authenticity and model-measured stylistic similarity, with post-edited text often perceived as representative of participants' personal style despite remaining detectable LLM stylistic traces.

Downloads

Next from ACL 2026

Paramanu: Compact and Competitive Monolingual Language Models for Low-Resource Morphologically Rich Indian Languages

Similar lecture

Layer-wise Swapping for Generalizable Multilingual Safety

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES