This digest collects 20 AI/ML papers posted to arXiv on 2026-03-26, spanning RAG, speech, multimodal reasoning, hardware optimization, evaluation, TTS, green AI, retraining policies, uncertainty, embodied AI, safety, intent representation, reasoning puzzles, federated learning, traffic, manipulation, knowledge graphs, quantization, biometrics, and neural network verification.
Key points
- WriteBack-RAG (2603.25737) — A framework that distills retrieval successes into compact knowledge units and re-indexes them with the original corpus. Plug-in compatible with any RAG pipeline; averages +2.14% across 4 RAG methods, 6 benchmarks, and 2 LLM backbones.
- WildASR (2603.25727) — Multilingual diagnostic benchmark built entirely from real human speech. Decomposes ASR robustness along environmental degradation, demographic shift, and linguistic diversity; reveals severe performance drops, poor cross-condition transfer, and hallucinations under degraded audio.
- RC2 (2603.25720) — Reinforcement learning framework that enforces cross-modal cycle consistency by forcing models to reverse-infer, switch modalities, and forward-reason back to the answer. Yields up to +7.6 percentage points in reasoning accuracy.
- Agent Factory (2603.25719) — Two-stage multi-agent pipeline for hardware design optimization with no hardware-specific training. Stage 1 splits kernels, optimizes each with pragmas, and solves an ILP for global assembly; Stage 2 launches N expert agents for cross-function optimization. Scaling from 1 to 10 agents gives an average 8.27× speedup across 12 kernels.
- Math Solving vs. Step-Level Assessment (2603.25633) — Uses PROCESSBENCH to show that LLMs evaluate their own correct solutions more accurately than incorrect ones, and that assessment is generally harder than direct problem solving.
- Voxtral TTS (2603.25551, Mistral AI) — Expressive multilingual TTS with 3-second voice cloning, combining autoregressive semantic-token generation with flow-matching acoustic decoding via the Voxtral Codec. Human evaluators prefer it over ElevenLabs Flash v2.5 in 68.4% of comparisons; weights released under CC BY-NC.
- EcoThink (2603.25498) — Energy-aware adaptive inference using a distilled router that skips chain-of-thought for factual queries and reserves deep reasoning for complex ones. Cuts inference energy by 40.4% on average (up to 81.9% on web retrieval) with no statistically significant quality loss across 9 benchmarks.
- Learning Debt (2603.25480) — Reframes model retraining as approximate Bayesian inference with computational constraints. Defines learning debt as the gap between continuously updated beliefs and a frozen deployed model; derives retraining thresholds from the loss function for auditable, evidence-based triggers.
- Cross-Model Perplexity / Entropy (2603.25450) — Training-free correctness indicator that measures a second verifier model's surprise when reading the answer. CMP achieves 0.75 AUROC on MMLU versus 0.59 for in-model entropy, targeting confident-but-wrong failures.
- Embodied Semantic Scene Graphs (2603.25415) — Modular navigation component for embodied SSG generation. Swapping policy optimization for modern optimization lifts SSG completeness by 21% relative to baseline with compact, fine-grained discrete action sets.
- Reasoning Safety (2603.25412) — Formalizes safety of the chain-of-thought itself, introducing a 9-class taxonomy of unsafe reasoning (input parsing, reasoning execution, process management errors). The proposed monitor reaches 84.88% step localization and 85.37% error-type accuracy on a 450-chain static benchmark.
- PPS: Prompt Protocol Specification (2603.25379) — 5W3H-based structured intent representation evaluated across 2160 outputs (3 languages × 4 conditions × 3 LLMs × 60 tasks). AI-expanded 5W3H prompts match hand-written ones for goal alignment; structured PPS reduces or reshapes cross-model variance and exposes a double-inflation bias in unstructured prompts.
- Integer Arithmetic Puzzle Difficulty (2603.25356) — Formalizes a number-game-inspired puzzle class, builds an exact dynamic-programming solver, and generates 3.4M+ instances. Difficulty is fully determined by a small set of structural features derivable from minimal witnesses.
- Agentic Trust Coordination for FL (2603.25334) — Lightweight server-side Agentic Trust Control Layer for federated learning in industrial networks that observes trust and system signals, interprets their evolution, and applies targeted trust adjustments when instability is detected.
- DRL for Mixed Traffic (2603.25328) — Uses TD3-based AV controllers trained on NGSIM. RL-controlled traffic raises road capacity by ~7.52% over fully human driving; RL AVs achieve ~28.98% better fuel efficiency at higher speeds while maintaining safety and comfort.
- Harmful AI Manipulation Framework (2603.25326) — Context-specific human-AI interaction study with 10,101 participants across public policy, finance, and health in the US, UK, and India. Models can produce manipulative behavior on prompt and induce belief/behavior change; effects vary by domain and do not transfer across regions.
- DAGverse (2603.25293) — Framework for constructing document-grounded semantic DAGs from scientific papers. Releases DAGverse-1, a dataset of 108 expert-validated semantic DAGs built via classification, reconstruction, semantic grounding, and verification.
- SliderQuant (2603.25284) — Layer-aware post-training quantization. Empirically shows shallow/deep layers (especially first/last) are more sensitive than middle layers, then applies adaptive sliding quantization. Outperforms existing PTQ on Llama/Llama2/Llama3/Qwen2.5, DeepSeek-R1 distilled, and large MoE models.
- Gait Foundation Model (2603.25283) — 3D skeleton gait foundation model from 3,414 deeply phenotyped adults. Embeddings predict age (r=0.69), BMI (r=0.90), and visceral adipose tissue area (r=0.82), and significantly forecast 1,980 of 3,210 phenotypes. Leg-dominant embeddings predict metabolic and frailty traits; trunk embeddings capture sleep and lifestyle.
- Distribution and Clusters Approximation (2603.25273) — Two new approximation methods—distribution and clusters—for the grid abstract domain used in probabilistic abstract interpretation of neural networks. Includes theoretical analysis and illustrative abstract transformers.
- Collection date: 2026-03-30
- Total papers: 20
- All arXiv links:
https://arxiv.org/abs/<id>as listed above