United States

Enterprise LLM agents can dramatically improve workplace productivity, but their core capability, retrieving and using internal context to act on a user’s behalf, also creates new risks for sensitive information leakage. 
We introduce **CI-Work**, a Contextual Integrity (CI)-grounded benchmark that simulates enterprise workflows across five information-flow directions and evaluates whether agents can convey *essential* content while withholding *sensitive* context in dense retrieval settings.
Our evaluation of frontier models reveals that privacy failures are prevalent (violation rates range from 15.8%-50.9%, with leakage reaching up to 26.7%) and uncovers a counterintuitive trade-off critical for industrial deployment: higher task utility often correlates with increased privacy violations.
Moreover, the massive scale of enterprise data and potential user behavior further amplify this vulnerability. Simply increasing model size or reasoning depth fails to address the problem. We conclude that safeguarding enterprise workflows requires a paradigm shift, moving beyond model-centric scaling toward context-centric architectures.

ACL 2026

CI-Work: Benchmarking Contextual Integrity in Enterprise LLM Agents

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

technical paper

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. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<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>

To access this event page, you need to log in with the **email address you registered with**. 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.

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.

Large Language Models (LLMs) have achieved remarkable performance in Machine Translation (MT), but deploying them at scale remains prohibitively expensive. 
A widely adopted remedy is the hybrid system paradigm, which balances cost and quality by serving most requests with a small model and selectively routing a fraction to a large model. 
However, existing routing strategies often rely on heuristics, external predictors, or absolute quality estimation, which fail to capture whether the large model actually provides a worthwhile improvement over the small one. 
In this paper, we formulate routing as a budget allocation problem and identify marginal gain, i.e., the large model’s improvement over the small model, as the optimal signal for budgeted decisions. 
Building on this, we propose \textbf{RouteLMT} (routing for LLM-based MT), an efficient in-model router that predicts this expected gain by probing the small translator’s prompt-token representation, without requiring external models or hypothesis decoding. 
Extensive experiments demonstrate that our RouteLMT outperforms heuristics, quality/difficulty estimation baselines, achieving a superior quality–budget Pareto frontier. 
Furthermore, we analyze regression risks and show that a simple guarded variant can mitigate severe quality losses.

RouteLMT: Learned Sample Routing for Hybrid LLM Translation Deployment

Automatic speech recognition (ASR) systems, despite low overall word error rates, produce residual lexical errors that disproportionately affect semantically critical tokens such as named entities, negations, and sentiment-bearing words. These errors are often structured, arising from phonetic similarity rather than random noise, making naive token-level correction insufficient.
We propose a structured ASR correction framework, that we call G-SPIN, that combines phonetic graph modeling with contextual language understanding. A graph neural network (GNN) first constructs acoustically plausible candidate neighborhoods for flagged tokens, explicitly restricting the correction search space to phonetic alternatives. A masked language model (MLM) then provides local contextual scoring, and an instruction-tuned large language model (LLM) performs final context-aware re-ranking over this compact candidate set. By decoupling structured phonetic reasoning from contextual semantic selection, our method avoids unconstrained generation while improving correction accuracy. The framework is lightweight, modular, and operates entirely at inference time.

Graph-Based Phonetic Error Correction of Noisy ASR

Large Reasoning Models (LRMs) have achieved remarkable success on complex tasks by leveraging long chain-of-thought (CoT) trajectories, yet they frequently exhibit overthinking on simple queries, resulting in significant token overhead and reduced inference efficiency. However, existing compression methods predominantly apply uniform length reduction or rely on coarse-grained difficulty estimation, often leading to performance degradation on difficult problems. To address this limitation, we propose Confidence-Adaptive Thinking (CAT), a framework that incorporates the model's intrinsic self-certainty signals as confidence into the preference optimization process, which autonomously modulates reasoning lengths based on problem difficulty. Experimental results show that CAT consistently outperforms state-of-the-art baselines on reasoning accuracy across multiple benchmarks on different base models. Our work enables LRMs to effectively compress confident responses while deliberating on uncertain ones, offering a potentially robust solution for balancing accuracy and latency in practical industrial scenarios.

CAT: Confidence-Adaptive Thinking for Efficient Reasoning of Large Reasoning Models

Prompt choice is crucial in adapting language models to text classification tasks, particularly under low-resource conditions. Manual prompt engineering is time-consuming, non-scalable, and brittle, while current auto-prompting techniques are still far from maturity. This paper presents a two-stage method for prompt learning of frozen language models, CRL-Prompt, based on soft prompt initialization followed by contrastive and reinforcement-based refinement. An experimental study demonstrates that our approach achieves consistent improvements in accuracy over baseline prompt tuning strategies, with gains of up to 2.2% while training fewer than 0.25% of model parameters.

CRL-Prompt: Contrastive and Reinforcement Learning for Soft Prompt Tuning for Text Classification

Pretokenization is a crucial, sequential pass in Byte-level BPE tokenizers, yet little work has been done to optimize it for edge-side inference. Our proposed new implementation, Peek2, serves as a drop-in replacement for cl100k-like pretokenizers used in GPT-3, LLaMa-3, and Qwen-2.5. After breaking down and analyzing the logic of the original cl100k pretokenizer, we introduced a new pretokenization algorithm with linear time complexity and constant, trivial memory usage, suited for edge scenarios. Test results show that it increases microbenchmarking throughput by up to $ 2.48\times $ and delivers a $ 1.14\times $ improvement in overall throughput across the entire Byte-level BPE en coding process, depending on the dataset, while providing identical results as the baseline Regex-based tokenizer.

Peek2: Regex-free Byte-level Byte-Pair Encoding Pretokenizer for LLM Inference on Edge Devices

Diffusion language models generate text by iteratively denoising all tokens in parallel, but when and where their hidden states encode whether the output will be functionally correct remains unknown. We present the first probing study of DLM internals, training linear classifiers on hidden states extracted at intermediate denoising steps to predict functional correctness. Across two models (LLaDA-8B, Dream-7B) and four tasks (JSON schema validation, math reasoning, code generation, science QA), we find that (1) correctness signal is already present at step 0, before any denoising occurs, (2) signal emergence is task-dependent, with structural tasks showing flat profiles and reasoning tasks showing gradual buildup, and (3) layer dynamics diverge across architectures. We further demonstrate that per-step probe confidence can identify instances likely to fail, enabling selective generation that avoids 36-98% of wasted compute depending on the task.

Probing Functional Correctness in Diffusion Language Models

Tokenizer design is a foundational but underexamined component of large language model (LLM) development. We introduce TokLens, an open-source toolkit for evaluating tokenizer quality across languages using six intrinsic metrics: fertility, characters per token, compression ratio, normalized sequence length, single-token retention rate, and cross-lingual parity. We evaluate 24 tokenizers from major LLM families across 15 typologically diverse languages and correlate intrinsic metrics with downstream performance on the Open LLM Leaderboard v2. Our analysis reveals stark cross-lingual disparities: GPT-2's tokenizer produces 56x more tokens per word in Japanese than in English, while newer tokenizers like Qwen2.5 and Gemma-2 reduce this gap to under 4x. No intrinsic metric predicts English benchmark performance after controlling for model size. However, when we extend the analysis to multilingual benchmarks (MMLU-ProX) using linear mixed-effects models that account for repeated measures within models, tokenizer metrics significantly predict per-language performance (STRR: $\beta = +5.7$, $z = 18.5$, $p < 0.001$). A controlled experiment with the Qwen2.5 family, which shares a single tokenizer across five model sizes, reveals that languages with higher single-token retention rate benefit more from scaling ($\rho = 0.91$, $p < 0.001$). These results suggest that tokenizer quality has a measurable impact on multilingual LLM performance, even when model architecture and training pipeline are held constant.

TokLens: A Multilingual Lens on Tokenizer Quality for LLMs

Large language models are increasingly used as evaluators, yet recent work shows that their judgments can be distorted by superficial properties of responses, including epistemic markers such as maybe, probably, and I think. Prior evidence has suggested that hedged but correct answers may be penalized by LLM judges, but it remains unclear whether this effect survives in multilingual and code-mixed settings once benchmark construction artifacts are controlled. We present PolyJudge-Uncertain, a controlled benchmark for multilingual judge robustness under hedging, built from 320 base question-answer items and expanded into 5,120 examples across English, Hindi, Hinglish, and Bengali, with balanced assertive versus hedged formulations and correct versus incorrect answers. We evaluate a reference-guided judge using both pointwise and pairwise protocols. In an initial pilot, we observe a large pointwise penalty against hedged answers. However, after naturalizing templates and introducing template-quality control, that pointwise effect largely disappears: pointwise accuracy reaches 99.8\% overall, with no meaningful assertive--hedged gap. The robust remaining effect emerges in pairwise comparison. When two answers are equally correct and differ only in style, the judge prefers the assertive response in 1,276 of 1,280 cases. These results refine the current narrative on uncertainty bias in LLM-as-a-judge. In our setting, the strongest stable failure mode is not pointwise misclassification of hedged truth, but pairwise over-preference for assertiveness among semantically equivalent outputs. More broadly, our findings show that multilingual judge-bias claims are highly sensitive to template naturalness and benchmark construction, making benchmark auditing a first-class methodological requirement for this line of work.

Confidence as a Tie-Breaker: Reassessing Multilingual Hedging Bias in LLM-as-a-Judge Evaluation

As large language models (LLMs) grow, their compute and memory demands become prohibitive for on-device deployment. Quantization is a crucial technique to shrink model footprint and accelerate inference, but aggressively low-bit weight-activation quantization schemes often sacrifice accuracy. Quantization Aware Training (QAT) is a commonly used paradigm to minimize quantization noise, but is extremely expensive to train and often unscalable to large models. We introduce PE-QAT, a parameter-efficient framework targeting per-channel 4-bit weight-activation quantization of LLMs, which aims to preserve model accuracy while significantly reducing resource requirements. The proposed method freezes the base model and trains lightweight LoRA adapters by fake quantizing the merged-weight model, enabling PE-QAT to scale efficiently unlike full QAT. We apply fake quantization with Straight-Through Estimators (STE) to the merged weights, allowing the adapters to explicitly compensate for quantization noise during training. One of the biggest challenges with quantizing activations alongside weights is addressing outliers that are orders of magnitude larger than other activations, which inflate quantization scales and suppress lower-magnitude values. To mitigate the impact of severe activation outliers, PE-QAT jointly learns per-channel smoothing factors and symmetric activation clipping thresholds. PE-QAT retains accuracy within 0.11 percentage points of the full-precision baseline on Llama-2-7B zero-shot tasks while training only 1.26% of total parameters.

PE-QAT: Parameter-Efficient Quantization Aware Training for Large Language Models

Evaluating the grammatical abilities of large language models (LLMs) is important for both NLP and linguistic theory. We investigate the ability of large language models (LLMs) to perform acceptability judgments in a forced-choice paradigm. We evaluate a subset of LLMs on 150 minimal sentence pairs sampled from Linguistic Inquiry and categorized using BLiMP linguistic phenomena. Our results show that while LLMs approximate human judgments, performance varies across models and phenomenon types, with less stable behavior on more complex phenomena. Prompting strategies have minimal impact.

Downloads

Next from ACL 2026

RouteLMT: Learned Sample Routing for Hybrid LLM Translation Deployment

Similar lecture

Can LLMs Help You at Work? A Sandbox for Evaluating LLM Agents in Enterprise Environments

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES