🗣️ Dialogue Systems

🧠 NeurIPS2025 · 8 paper notes

📌 Same area in other venues: 📷 CVPR2026 (1) · 🔬 ICLR2026 (10) · 💬 ACL2026 (26) · 🧪 ICML2026 (5) · 🤖 AAAI2026 (5)

🔥 Top topics: LLM ×2

AC-LoRA: (Almost) Training-Free Access Control-Aware Multi-Modal LLMs

AC-LoRA is an end-to-end system that trains independent LoRA adapters for datasets with different permission levels. At inference time, it dynamically retrieves and training-freely merges multiple LoRA outputs based on cosine similarity and user permissions, achieving strong information isolation while matching or surpassing SOTA LoRA mixture methods in response quality.

Agentic Persona Control and Task State Tracking for Realistic User Simulation

A three-agent collaborative framework for realistic user simulation is proposed, comprising a User Agent (coordination), a State Tracking Agent (structured task state), and a Message Attributes Generation Agent (behavior attribute control conditioned on persona and state). On a restaurant ordering scenario, the framework achieves a 102.6% improvement in composite realism score (CRRS), +19.9% in persona adherence, and +284.5% in behavioral variability. A core finding is that behavior control without state awareness yields BVS = 0 (completely rigid behavior).

Bridging Human and LLM Judgments: Understanding and Narrowing the Gap

This paper proposes Bridge, a statistical framework that models the latent relationship between human and LLM judgments via ordinal logistic regression. With a small number of human labels, Bridge improves the calibration and alignment of LLM judgments while supporting formal statistical hypothesis testing for systematic biases.

HyGen: Efficient LLM Serving via Elastic Online-Offline Request Co-location

This paper proposes HyGen, an interference-aware LLM inference system that achieves elastic co-location of online and offline workloads through an accurate batch latency predictor, an SLO-aware performance profiler, and a prefix-sharing-maximization scheduling strategy, delivering 3.87–5.84× throughput gains while strictly guaranteeing SLO compliance.

KL Penalty Control via Perturbation for Direct Preference Optimization

This paper proposes ε-DPO, which achieves instance-level adaptive KL penalty control by monitoring the monotonicity of logits—used as preference model outputs—under small perturbations of \(\beta\) during training. The method incurs no additional computational overhead and significantly outperforms DPO and most direct alignment algorithms, achieving a 46.4% LC win rate on AlpacaEval 2 (vs. 40.3% for DPO).

LatentGuard: Controllable Latent Steering for Robust Refusal of Attacks and Reliable Response Generation

This paper proposes LatentGuard, a three-stage framework that combines behavior-level alignment fine-tuning, structured VAE-supervised latent space modeling, and latent-space dimensional manipulation to achieve interpretable and controllable regulation of LLM refusal behavior — robustly defending against adversarial attacks while preserving responsiveness to benign queries.

Less is More: Local Intrinsic Dimensions of Contextual Language Models

This paper proposes using the Local Intrinsic Dimension (LID) of contextual token embeddings as an unsupervised signal for monitoring LLM training dynamics — a decrease in LID indicates improved generalization, while an increase signals overfitting. The utility of this geometric signal is validated on tasks including dialogue state tracking, grokking, and sentiment recognition.

SciArena: An Open Evaluation Platform for Non-Verifiable Scientific Literature-Grounded Tasks

SciArena is a community-driven open evaluation platform for scientific literature tasks. It adopts a Chatbot Arena-style human preference voting paradigm to rank 47 foundation models, collecting over 20,000 votes, and releases SciArena-Eval as a meta-benchmark for assessing the ability of automated evaluation systems to judge answer quality on literature-grounded tasks.