📚 arXiv Daily AI/ML Paper Digest (2026-03-30): 20 Selected Papers

📅 Collection date: 2026-03-30 📊 Paper count: 20 papers (all published 2026-03-26)

---

1. WriteBack-RAG: Trainable Knowledge Bases for Retrieval-Augmented Generation

Authors: Yuxing Lu et al. — arXiv:2603.25737

RAG knowledge bases are typically assembled once and never revised, yet facts needed by queries are often scattered across documents. WriteBack-RAG uses annotated examples to identify successful retrieval locations, isolates relevant documents, and distills them into compact knowledge units indexed alongside the original corpus. The method can be preprocessed offline and combined with any RAG pipeline. Evaluated across four RAG methods, six benchmarks, and two LLM backbones, it delivers an average improvement of 2.14%.

2. WildASR: A Multilingual Diagnostic Benchmark for ASR Robustness in the Wild

Authors: Geeyang Tay, Wentao Ma et al. — arXiv:2603.25727

While ASR systems approach human-level accuracy on curated benchmarks, they still fail in real-world voice agents. WildASR is a multilingual diagnostic benchmark built entirely from real human speech, decomposing ASR robustness along three axes: environmental degradation, demographic shifts, and linguistic diversity. Evaluations reveal severe performance drops, robustness that does not transfer across languages or conditions, and models frequently hallucinating content on degraded input.

3. RC2: Reinforcement Learning with Cross-Modal Cycle Consistency

Authors: Zirui Zhang et al. — arXiv:2603.25720

Current multimodal models often produce contradictory predictions for visual and textual representations of the same concept. RC2 is a reinforcement learning framework that enforces cross-modal cycle consistency: the model must perform backward reasoning, switch modality, and reliably reconstruct the answer via forward reasoning, yielding dense label-free rewards. Optimizing this structure improves reasoning accuracy by up to 7.6 percentage points.

4. Agent Factory: Autonomous Hardware Design Optimization with Multi-Agent Systems

Authors: Abhishek Bhandwaldar et al. — arXiv:2603.25719

An empirical study of general-purpose coding agents (no hardware-specific training) optimizing hardware designs from high-level algorithmic specifications. The two-stage pipeline decomposes designs into sub-kernels, independently optimizes each via pragma and code-level transformations, formulates an integer linear program (ILP) to assemble a global configuration, then launches N expert agents to explore cross-function optimization. On 12 kernels, scaling from 1 to 10 agents yields an average 8.27x speedup.

5. On the Relationship Between Math Problem Solving and Step-Level Assessment in LLMs

Authors: Liang Zhang et al. — arXiv:2603.25633

Using the ProcessBench benchmark, this study examines how math problem-solving ability relates to step-level assessment performance. A consistent within-model pattern emerges: assessment accuracy is significantly higher on problems the model itself solves correctly. Assessment is also harder than direct problem solving, especially on solutions containing errors.

6. Voxtral TTS: Expressive Multilingual Text-to-Speech with 3-Second Voice Cloning

Authors: Alexander H. Liu et al. (Mistral AI) — arXiv:2603.25551

Voxtral TTS generates natural speech from just 3 seconds of reference audio. It uses a hybrid architecture combining autoregressive generation of semantic speech tokens with flow matching for acoustic tokens, encoded/decoded via the Voxtral Codec (a speech tokenizer trained from scratch). In human evaluations it is preferred for naturalness and expressiveness, with a 68.4% win rate over ElevenLabs Flash v2.5. Model weights are released under a CC BY-NC license.

7. EcoThink: Energy-Aware Adaptive Inference for Sustainable Generative AI Agents

Authors: Zhixiang Lu et al. — arXiv:2603.25498

Current paradigms indiscriminately apply compute-intensive chain-of-thought strategies to billions of daily queries, causing LLM overthinking. EcoThink uses a lightweight distillation-based router to dynamically assess query complexity, skipping unnecessary reasoning for factual retrieval while preserving deep computation for complex logic. Across nine benchmarks it reduces inference energy by 40.4% on average (up to 81.9% for web knowledge retrieval) with no statistically significant performance loss.

8. Learning Debt: A Decision-Theoretic Framework for Model Retraining

Author: Harrison Katz — arXiv:2603.25480

Retraining can be better understood as approximate Bayesian inference under computational constraints. The gap between continuously updated belief states and the frozen deployed model is "learning debt," and retraining decisions become cost-minimization problems with thresholds derived from the loss function. The framework replaces calendar-based schedules with evidence-based triggers, making governance auditable.

9. Cross-Model Perplexity: A Training-Free Correctness Indicator for Language Models

Authors: Matt Gorbett et al. — arXiv:2603.25450

Existing error-detection methods rely on a model's own uncertainty, which fails badly on the most dangerous failure mode: confident mistakes. This work introduces cross-model disagreement as a simple, training-free correctness signal — a single forward pass of a verifier model measures how surprised it is reading the answer. Instantiated as cross-model perplexity (CMP) and cross-model entropy (CME), CMP reaches 0.75 average AUROC on MMLU versus 0.59 for within-model entropy baselines.

10. Embodied Semantic Scene Graph Generation with Modern Optimization

Authors: Roman Küble et al. — arXiv:2603.25415

This work presents a modular navigation component for embodied semantic scene graph (SSG) generation, replacing policy optimization with modern optimization methods, comparing single-head policies over atomic actions vs. decomposed multi-head policies over action components. Simply swapping the optimization algorithm improves SSG completeness by 21% relative to baselines.

11. Reasoning Safety: A Taxonomy and Monitor for Unsafe Reasoning in Large Reasoning Models

Authors: Xunguang Wang et al. — arXiv:2603.25412

Reasoning safety is identified as a co-equal safety dimension: reasoning trajectories must be logically consistent, computationally efficient, and resistant to adversarial manipulation. The paper formally defines reasoning safety, introduces a nine-category taxonomy of unsafe reasoning behaviors (input parsing, reasoning execution, and process management errors), and proposes a reasoning safety monitor achieving 84.88% step-level localization accuracy and 85.37% error-type classification accuracy on a static benchmark of 450 reasoning chains.

12. PPS: Cross-Language and Cross-Model Study of Structured Intent Representation

Authors: Gang Peng et al. — arXiv:2603.25379

Across 2,160 model outputs (3 languages × 4 conditions × 3 LLMs × 60 tasks), the study finds that AI-expanded 5W3H prompts (Prompt Protocol Specification) show no statistically significant difference in goal alignment versus manually written 5W3H prompts — with users inputting only one sentence. Structured PPS conditions reduce or reshape cross-model output variance, while unstructured prompts exhibit a systematic double-inflation bias: artificially high composite scores and artificially low apparent cross-model variance.

13. Difficulty in Integer Arithmetic Puzzles: A Structural Analysis

Authors: Yunus Zeytuncu et al. — arXiv:2603.25356

Using an exact dynamic-programming solver, the authors build a dataset of over 3.4 million instances with difficulty defined by the minimum number of operations needed to reach a target. Analysis shows difficulty is fully determined by a small set of interpretable structural properties derived from exact witnesses, with the number of input values used in the minimal construction serving as a minimal sufficient statistic.

14. Agentic Trust Coordination for Federated Learning in Industrial Networks

Authors: Paul Shepherd et al. — arXiv:2603.25334

Federated learning reliability in industrial networks suffers from inconsistent client behavior, noisy sensing conditions, and faulty or adversarial updates. The proposed Agentic Trust Control Layer runs as a server-side control loop, observing trust and system-level signals, interpreting their evolution over time, and applying targeted trust adjustments when instability is detected.

15. Deep Reinforcement Learning for Mixed Traffic: Macroscopic Analysis and Fuel Efficiency

Authors: Pankaj Kumar et al. — arXiv:2603.25328

Using TD3 for autonomous vehicle control trained on the NGSIM highway dataset, the study analyzes macroscopic traffic flow characteristics of DRL-based control in mixed traffic. Road capacity increases by ~7.52% from fully human-driven to fully RL-controlled traffic, and RL-based AVs improve average fuel efficiency by ~28.98% at higher speeds.

16. Evaluating Harmful AI Manipulation: A Framework for Context-Specific Human-AI Interaction Studies

Authors: Canfer Akbulut et al. — arXiv:2603.25326

Evaluated with 10,101 participants across three AI use domains (public policy, finance, health) and three regions (US, UK, India), the tested models can produce manipulative behavior when prompted and can induce belief and behavior changes in participants in experimental settings. Manipulation varies across domains, and results from one geographic region may not generalize to others.

17. DAGverse: Constructing Document-Grounded Semantic DAGs from Scientific Papers

Authors: Shu Wan et al. — arXiv:2603.25293

DAGverse builds document-grounded semantic DAGs from scientific papers containing explicit DAG figures as natural supervision. Its core component, DAGverse-Pipeline, is a semi-automatic system producing high-precision semantic DAG examples via graph classification, graph reconstruction, semantic grounding, and verification. The released DAGverse-1 dataset contains 108 expert-verified semantic DAGs.

18. SliderQuant: Sliding-Layer Quantization for Large Language Models

Authors: Anbang Yao et al. — arXiv:2603.25284

Empirical study of per-layer quantization sensitivity shows: (1) shallow/deep layers are generally more quantization-sensitive than middle layers; (2) within shallow/deep layers, the first/last layers are most sensitive. SliderQuant is a post-training quantization framework built on adaptive sliding quantization, outperforming existing PTQ methods on Llama/Llama2/Llama3/Qwen2.5 families, DeepSeek-R1 distill models, and large MoE models.

19. A Gait Foundation Model for Multi-System Biomarker Discovery

Authors: Adam Gabet et al. — arXiv:2603.25283

A gait foundation model for 3D skeletal motion, trained on 3,414 deeply phenotyped adults recorded via depth cameras across five movement tasks. Learned embeddings outperform engineered features, predicting age (Pearson r=0.69), BMI (r=0.90), and visceral adipose tissue area (r=0.82). Embeddings significantly predict 1,980 of 3,210 phenotype targets; after adjusting for age, BMI, VAT, and height, gait provides independent gains for all 18 body systems in males and 17 of 18 in females. Legs dominate metabolic and frailty prediction, while the torso encodes sleep and lifestyle phenotypes.

20. Distribution and Clusters Approximation for Neural Network Abstract Interpretation

Authors: Zhuofan Zhang et al. — arXiv:2603.25273

Within a probabilistic abstract interpretation framework for neural network analysis, this paper introduces two new approximation methods — distribution approximation and clusters approximation — describing their theoretical workings and demonstrating the corresponding abstract transformers via simple illustrative examples.

---

*All arXiv IDs and links are preserved as cited in the original post.*