United States

Role-playing agents (RPAs) require balancing multiple objectives, such as instruction following, persona consistency, and stylistic fidelity, which are not always perfectly aligned across different dimensions. While prior work has primarily relied on supervised fine-tuning or reinforcement learning with scalarized rewards, these approaches do not explicitly address the coordination of multiple reward dimensions during optimization. We present **MOA** (**M**ulti-**O**bjective **A**lignment), a reinforcement-learning framework that enables multi-dimensional, fine-grained rubric optimization for general RPAs. MOA introduces a novel multi-objective optimization strategy that trains simultaneously on multiple fine-grained rubrics to boost optimization performance. Besides, to address the issues of model output diversity and quality, we have also employed thought-augmented rollout with off-policy guidance. Experiments on PersonaGym and RoleMRC show that MOA consistently improves multi-dimensional role-playing performance over supervised and standard RL baselines. Under identical evaluation protocols, an 8B model trained with MOA reaches performance competitive with strong closed-source models across multiple evaluation dimensions. These results suggest that MOA provides a practical framework for training more capable general-purpose role-playing agents.

ACL 2026

MOA: Multi-Objective Alignment for Role-Playing Agents

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.

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.

technical paper

While LLMs are powerful embedding backbones, their application in training-free settings faces two structural challenges: causal attention restricts early tokens from accessing subsequent context, and the next-token prediction objective biases representations toward generation rather than semantic compression. To address these limitations, we propose KV-Embedding, a framework that activates the latent representation power of frozen LLMs. Our method leverages the observation that the key-value (KV) states of the final token at each layer encode a compressed view of the sequence. By re-routing these states as a prepended prefix, we enable all tokens to access sequence-level context within a single forward pass. To ensure model-agnostic applicability, we introduce an automated layer selection strategy based on intrinsic dimensionality. Evaluations on MTEB across Qwen, Mistral, and Llama backbones show that KV-Embedding outperforms existing training-free baselines by up to 10%, while maintaining robust performance on sequences up to 4,096 tokens. These results demonstrate that internal state manipulation offers an efficient alternative to input modification, and we hope this work encourages further exploration of LLM internals for representation learning.

KV-Embedding: Training-free Text Embedding via Internal KV Re-routing in Decoder-only LLMs

Chain-of-thought (CoT) reasoning has emerged as a crucial paradigm for enhancing large language model (LLM) performance on multi-step reasoning tasks.
However, the internal mechanisms by which LLMs invoke knowledge and propagate information across different steps of the CoT are poorly understood.
To fill this gap, we propose a multi-stage probing framework that enforces structured reasoning with three explicit stages: keyword extraction, theorem generation, and computation execution.
The framework integrates attention knockout to trace cross-layer information flow and theorem probing to examine how specific contents are encoded within representations.
To enable controlled and stage-aligned analysis, we construct a structured CoT dataset that covers the mathematics and physics domains. Experiments on four instruction-tuned LLMs reveal distinct stage-specific patterns.
First, keyword information is progressively aggregated into the final token in later layers.
Second, theorem semantics are encoded in the mid-to-late layers and undergo two stages of propagation.
Finally, parameter substitution is achieved through joint extraction by the final token and other tokens.
The first parameter predominantly relies on the final token, whereas later parameters increasingly depend on information extracted by other tokens.
Overall, our findings shed light on the neural implementation of CoT reasoning and provide actionable insights for developing more interpretable and reasoning-capable LLMs.
We further evaluate a free-form prompting setting without labeled fields and observe consistent qualitative trends.

Exploring Layer Activation Dynamic of CoT via Knowledge Probe

Unified Information Extraction (UIE) aims to handle heterogeneous IE tasks within a single framework, but existing methods often suffer from inconsistent schema representation, implicitly intermediate reasoning and full-parameter adaptation, which limit generalization, interpretability and parameter efficiency. To address these issues, we propose UC-UIE (Universal Capabilities-based Unified Information Extractor), a unified framework based on Large Language Model (LLM), which introduces a unified frame-and-slots schema for IE tasks and explicitly decomposes IE reasoning into three universal capabilities: judging, locating, and associating. Furthermore, UC-UIE adopts a Low-Rank Adaptation (LoRA) based hierarchical Mixture-of-Experts (MoE) adapter to fine-tune LLMs for IE tasks, which explicitly models these three capabilities in a task-driven way while ensuring parameter efficiency. With only 1.24% trainable parameters, UC-UIE outperforms full-parameter tuning methods, showing excellent parameter efficiency. Zero-shot evaluation reveals its strong generalization ability to unseen domains and schemas, benefiting from unified schema representation and explicit capability decomposition. Further experiments validate that the hierarchical MoE adapter learns capability specialization and composition, which enhances both UIE performance and interpretability.

Capability Decomposition for Unified Information Extraction via Hierarchical Mixture-of-Experts

The widespread utilization of language models in modern applications is inconceivable without Parameter-Efficient Fine-Tuning techniques, such as low-rank adaptation ($\texttt{LoRA}$), which adds trainable adapters to selected layers. Although $\texttt{LoRA}$ may obtain accurate solutions, it requires significant memory to train large models and intuition on which layers to add adapters. In this paper, we propose a novel method, $\texttt{WeightLoRA}$, which overcomes this issue by adaptive selection of the most critical $\texttt{LoRA}$ heads throughout the optimization process. As a result, we can significantly reduce the number of trainable parameters while maintaining the capability to obtain consistent or even superior metric values. We conduct experiments for a series of competitive benchmarks and DeBERTa, BART, Llama and Qwen models, comparing our method with different adaptive approaches. The experimental results demonstrate the efficacy of $\texttt{WeightLoRA}$ and the superior performance of $\texttt{WeightLoRA+}$ in almost all cases. The source code is available at https://github.com/brain-lab-research/WLoRA

WeightLoRA: Keep Only Necessary Adapters

LLM-based agents are increasingly deployed to autonomously solve complex tasks, raising urgent needs for IP protection and regulatory provenance. While content watermarking effectively attributes LLM-generated outputs, it fails to directly identify the high-level planning behaviors (e.g., tool and subgoal choices) that govern multi-step execution. Critically, watermarking at the planning-behavior layer faces unique challenges: minor distributional deviations in decision-making can compound during long-term agent operation, degrading utility, and many agents operate as black boxes that are difficult to intervene in directly. To bridge this gap, we propose AgentMark, a behavioral watermarking framework that embeds multi-bit identifiers into planning decisions while preserving utility. It operates by eliciting an explicit behavior distribution from the agent and applying distribution-preserving conditional sampling, enabling deployment under black-box APIs while remaining compatible with action-layer content watermarking. Experiments across embodied, tool-use, and social environments demonstrate practical multi-bit capacity, robust recovery from partial logs, and utility preservation. Code is available at https://github.com/Tooooa/AgentMark.

AgentMark: Utility-Preserving Behavioral Watermarking for Agents

Molecular graph learning benefits from positional signals that capture both local neighborhoods and global topology. Two widely used families are spectral encodings derived from Laplacian or diffusion operators and anchor-based distance encodings built from shortest-path information, but the relationship between them is still not well understood. In this paper, we study when anchor-based distance encodings can approximate diffusion geometry. Under a random r-regular graph model, we derive an explicit trilateration map that reconstructs truncated diffusion coordinates from transformed anchor distances and anchor spectral positions, together with pointwise and Frobenius-gap guarantees. On DrugBank with a shared GNP-based DDI backbone, anchor-distance Nyström accurately recovers diffusion geometry, and both DE and LapPE outperform models without positional encodings, with LapPE showing slightly more consistent performance.

Bridging Distance and Spectral Positional Encodings via Anchor-Based Diffusion Geometry Approximation

The supervised fine-tuning (SFT) stage is crucial for multimodal large language models (MLLMs), yet a comprehensive scaling law to guide the optimal model-data configuration remains lacking. In this paper, we make an initial attempt to address this gap. First, we theoretically demonstrate that directly computing the optimal computation frontier for MLLM-SFT, as we can for traditional LLMs, is a challenging task. This complexity arises because MLLM-SFT is influenced by a broader range of factors, including model size, LLM pre-training tokens, and MLLM SFT tokens. To tackle this issue, we propose two scaling laws based on LLM paradigms: one applicable when training data volumes are well defined by researchers, and another for cases where models are sourced from open communities with unknown training data. Through theoretical modeling and approximations, we provide researchers with valuable recommendations for optimal resource allocation. Furthermore, we establish a strong correlation ($ R^2 = 0.98$) between training loss and downstream performance, enabling accurate performance estimation without the need for exhaustive benchmarking. To validate our scaling laws, we construct a testbed of 60 models ranging from 50 million to 8 billion parameters, totaling 1,560 checkpoints. Each checkpoint is evaluated on than 10 MLLM benchmarks, ensuring robust fitting of our formulations.

Scaling Law for Multimodal Large Language Model Supervised Fine-Tuning

The most recent research uses reinforcement learning (RL) to post-train Multi-modal Large Language Models (MLLMs) such that these models are able to iteratively call search engines to dynamically access external knowledge when handling complex Visual Question Answering (VQA) tasks. However, existing methods face two major limitations in effectiveness and efficiency: i) For effectiveness, the objective of these methods, which only considers the correctness of the generated final response, overlooks the quality of intermediate search results, thus leading to suboptimal search strategies. ii) For efficiency, existing methods often unnecessarily invoke search calls during reasoning, making the inference inefficient. To address these issues, we propose \sys, a customized dual-objective reinforcement learning framework to improve the search strategies of MLLMs, enhancing their search quality yet minimizing search frequency. The key ideas include (1) a reward function that promotes correct reasoning trajectories with fewer search calls; and (2) dual optimization objectives that jointly optimize search quality and answer correctness. Extensive experiments on 3 real-world datasets demonstrate that DORA outperforms state-of-the-art methods, achieving up to 8.4% higher accuracy while reducing the number of search calls by 9.7%.

DORA: A Dual-Objective Reinforcement Learning Framework for Effective and Efficient Multimodal Agentic Search

Long-form question answering (LFQA) requires open-ended long-form responses that synthesize coherent, factually grounded content from multi-source evidence. This makes reinforcement learning (RL) reward design critical. The reward must be verifiable for faithful grounding and stable optimization. However, many standard rewards assume a unique target with an exact-match notion of correctness, which fits short-form QA and math but breaks in LFQA. As a result, current RAG systems still lack verifiable reward mechanisms, yielding unstable feedback signals and suboptimal optimization outcomes. We propose RioRAG, a framework for reinforced verifiable informativeness optimization. First, it defines informativeness as a measurable and externally verifiable objective for RL. Second, RioRAG uses nugget-centric verification with cross-source checks to enable self-evolution of smaller LLMs and to provide denser, action-discriminative rewards that mitigate reward sparsity and stabilize optimization. This formulation avoids handcrafted supervision for the policy model and strong teacher-model distillation, relying instead on externally verifiable feedback. Experiments on LongFact and RAGChecker show that RioRAG achieves higher factual recall and faithfulness, establishing verifiable reward modeling as a foundation for trustworthy long-form RAG. Our codes are available at https://github.com/RUCAIBox/RioRAG.

Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation

Existing long-term personalized dialogue systems struggle to reconcile unbounded interaction streams with finite context constraints, often succumbing to memory noise accumulation, reasoning degradation, and persona inconsistency. To address these challenges, this paper proposes Inside Out, a framework that utilizes a globally maintained PersonaTree as the carrier of long-term user profiling. By constraining the trunk with an initial schema and updating the branches and leaves, PersonaTree enables controllable growth, achieving memory compression while preserving consistency. Moreover, we train a lightweight MemListener via reinforcement learning with process-based rewards to produce structured, executable, and interpretable {ADD, UPDATE, DELETE, NO_OP} operations, thereby supporting the dynamic evolution of the personalized tree. During response generation, PersonaTree is directly leveraged to enhance outputs in latency-sensitive scenarios; when users require more details, the agentic mode is triggered to introduce details on-demand under the constraints of the PersonaTree. Experiments show that PersonaTree outperforms full-text concatenation and various personalized memory systems in suppressing contextual noise and maintaining persona consistency. Notably, the small MemListener model achieves memory-operation decision performance comparable to, or even surpassing, powerful reasoning models such as DeepSeek-R1-0528 and Gemini-3-Pro.

Downloads

Next from ACL 2026

KV-Embedding: Training-free Text Embedding via Internal KV Re-routing in Decoder-only LLMs

Similar lecture

MoA: Heterogeneous Mixture of Adapters for Parameter-Efficient Fine-Tuning of Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES