User Model Extraction, Multimodal and Statistical Alignment, and Efficient LLM Serving

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering user-model interpretability and safety in LLMs, efficient multimodal and educational serving systems, clinical time-series and autism recognition, nonparametric in-context learning theory, statistical alignment of generative outputs, and language-model-driven network traffic control.

1. User Model Extraction via Belief Self-Distillation

Authors: Ali Holmov, Yiran Huang, Kirill Bykov, Zeynep Akata | Categories: cs.LG, cs.CL Link: arxiv.org/abs/2609.31603

The paper introduces Belief Self-Distillation (BSD), a unified read-write framework that learns a compact user representation from a frozen LLM’s own conversations, making implicit user beliefs both decodable and causally writable. Across multiple model families, BSD recovers beliefs faithfully and enables stronger interventions than hidden-state steering, showing that refusal depends on inferred user intent rather than the request alone. It also finds that independently trained LLMs converge on a shared geometry for representing users.

Takeaway: A rare interpretability result with direct safety implications — models condition their refusals on who they think you are, and that belief can now be measured and edited.

2. DeepEdu-v1: Efficient and Scalable Agentic LLMs for Vietnamese Education

Authors: Quang Nguyen, Hieu Nguyen, Hien Hoang, Toan Pham, Cong Tran et al. | Categories: cs.AI Link: arxiv.org/abs/2609.31568

DeepEdu-v1 is an AI tutoring system for Vietnamese education built on SCALE, combining a long-context inference engine that amortizes token selection to cluster granularity and a self-improving agentic layer that curates a verified playbook instead of fine-tuning. In deployment it achieves nearly 2× TTFT speedup over standard vLLM serving and lifts agentic accuracy from 70.0% to 79.5% on complex tasks, with the strongest gains in financial-reasoning and interactive-agent benchmarks.

Takeaway: A compelling case study in sovereign, curriculum-aligned AI tutoring — addressing data-localization law, KV-cache bottlenecks on consumer GPUs, and hallucination on region-specific content all at once.

3. ViSTA: A Simple Bridge Extends Visual Alignment to Clinical Time-Series Understanding in Multimodal LLMs

Authors: Junyi Gao, Yu Shi, Pingzhao Hu, Ewen M Harrison | Categories: cs.CL, cs.AI Link: arxiv.org/abs/2609.31448

ViSTA is a compact adapter that injects irregular numerical clinical measurements into a pretrained vision-language model’s chart representations by learning corrections to visual tokens while leaving all pretrained parameters frozen. On MIMIC-IV, it achieves the highest mean scores across all four metrics for acute kidney injury and mortality prediction among compared adaptations at 2–9B parameters, with the 2B model reaching 0.7376 AUROC using just 0.516M trainable parameters.

Takeaway: Elegant parameter efficiency — reusing a vision-language model’s chart pathway for clinical time series gets a 2B model within striking distance of far larger reasoning-heavy systems.

4. UniAR: A Unified Framework for Autism Recognition Enhanced by Multi-View Prompt Learning

Authors: Lei Xin, Zeheng Wang, Jiayin Zhu, Shihong Huang, Fanhu Zeng et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2609.31298

UniAR tackles ASD recognition under scarce diagnostic text by using a large multimodal model to generate hierarchical word-, phrase-, and sentence-level descriptions, then aligning them with visual evidence via a Mixture-of-Experts Multi-Scale Alignment Module. Across four MRI and facial benchmarks it achieves 75.9% and 91.6% average accuracy respectively, improving over baselines by 1.5 and 1.2 percentage points.

Takeaway: A practical answer to semantic scarcity in clinical screening — synthetic multi-granularity prompts plus expert alignment yield interpretable gains without paired reports.

5. Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers

Authors: Jaehee Seo, Jisu Kim | Categories: stat.ML, cs.LG, math.ST Link: arxiv.org/abs/2609.31458

The authors study in-context learning under unknown local geometry modeled as sample-size-dependent mixtures of manifolds with heterogeneous dimensions, smoothness, and sampling masses. They establish a minimax lower bound capturing aggregate component difficulty, construct a matching oracle tangent local-polynomial estimator, and connect it to a two-stage softmax transformer with a geometric preconditioner and chartwise solvers that achieves the rate with logarithmic depth and polynomial size.

Takeaway: Moves ICL theory beyond Euclidean and single-manifold assumptions — an important step toward explaining why transformers handle geometrically heterogeneous real-world data so well.

6. Statistical attribute alignment for black-box generative AI via output post-processing

Authors: Kevin Jiang, Morgane Austern, Edgar Dobriban, Jason M. Klusowski | Categories: stat.ME, cs.AI, cs.LG, math.ST Link: arxiv.org/abs/2609.31607

The paper studies aligning the attribute distribution of black-box generative AI outputs with a user-specified target, motivated by fairness and synthetic data representation goals. The authors develop algorithms for exact and approximate alignment that minimize expected queries to the generator, prove optimality as the number of requested outputs grows, and show gains on text-to-image and geocoded persona generation tasks.

Takeaway: A theoretically grounded complement to prompting-based interventions — useful for anyone who needs distributional guarantees from models they can only query, not modify.

7. EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models

Authors: Kunxiong Zhu, Zhihao Shu, Hangyu Zheng, Minghai Qin, Miao Yin et al. | Categories: cs.DC, cs.LG, cs.PF Link: arxiv.org/abs/2609.31551

EAServe repositions the Encode phase as the control point of the Encode-Prefill-Decode pipeline for multimodal LLM serving, with a runtime managing load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning, plus a Hybrid Auto Selection layer for configuration search. Evaluated on image, video, and audio MLLMs, it delivers up to 4.3× and 1.7× higher goodput than NVIDIA Dynamo and vLLM under identical SLOs.

Takeaway: Identifies the structural resource imbalance that existing EPD frameworks ignore — the encode GPU starving downstream stages — and turns it into a control knob.

8. Intent2Tc: Automated Intent-to-Traffic Control Translation with Language Models

Authors: Andrea Masini, Sudipta Acharya, Paolo Bellavista, Luca Foschini, Burak Kantarci | Categories: cs.NI, cs.AI, cs.CL Link: arxiv.org/abs/2609.31397

Intent2Tc is a closed-loop LLM-driven framework that translates business-level traffic-shaping intents into declarative sub-intents and validated, executable Linux tc configurations, integrating an AQM-based digital twin, automated metadata extraction, critique-driven refinement, and RAG-based knowledge reuse. On 100 RFC 9315-compliant intents, Claude Sonnet-4.6 reaches 0.98 semantic similarity, 1.0 semantic unit coverage, and 0.045 normalized edit distance, while RAG lets compact models like Phi-4-mini approach much larger models.

Takeaway: A concrete demonstration that intent-based networking can close the loop from business language to deployable configs — with RAG doing the heavy lifting for smaller models.

This content was generated with AI assistance. Paper information sourced from arXiv.