Resource-efficient learning under limited compute, memory, and energy, including on-device LLMs, test-time adaptation and scaling, efficient LLMs, and embodied AI:
-
TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices
NeurIPS 2024
Uses early-exit ensembles for test-time adaptation on edge devices, cutting compute while adapting to distribution shift at inference time.
-
Edge-Cloud Collaborative Speech Emotion Captioning via Token-Level Speculative Decoding in Audio-Language Models
Interspeech 2026
Token-level speculative decoding across edge and cloud for speech emotion captioning, balancing on-device cost with richer affective grounding while limiting raw-audio exposure.
-
E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models
NeurIPS 2025
Backpropagation-free test-time adaptation for speech foundation models under acoustic shifts such as noise and accents, reducing memory cost at inference.
-
Position: Human-Robot Interaction Demands a Shift From Static Privacy Controls to Dynamic Learning
NeurIPS 2025, LAW
Argues that embodied agents need dynamic privacy learning grounded in Contextual Integrity, since static controls fail when agents infer sensitive conclusions from ambient data.
-
XTransfer: Modality-Agnostic Few-Shot Model Transfer for Human Sensing at the Edge
ICML 2026
Few-shot, modality-agnostic model transfer for human sensing on the edge when labelled sensor data are scarce and on-device compute is limited.
-
AdaNODEs: Test Time Adaptation for Time Series Forecasting Using Neural ODEs
ICASSP 2026
Source-free test-time adaptation for time-series forecasting using Neural ODE dynamics, targeting distribution shift at inference without labelled target data.
-
LiteFat: Lightweight Spatio-Temporal Graph Learning for Real-Time Driver Fatigue Detection
IROS 2025
Lightweight spatio-temporal graph learning for real-time driver fatigue detection on resource-limited embedded platforms where high-latency deep models are impractical.
-
LifeLearner: Hardware-Aware Meta Continual Learning System for Embedded Computing Platforms
ACM SenSys 2023
Hardware-aware meta continual learning that reduces memory and latency for on-the-fly personalization on resource-constrained embedded platforms.
-
UR2M: Uncertainty and Resource-Aware Event Detection on Microcontrollers
IEEE PerCom 2024
Uncertainty-aware event detection on microcontrollers that balances prediction reliability with tight compute and memory budgets at the edge.
-
PROS: an efficient pattern-driven compressive sensing framework for low-power biopotential-based wearables with on-chip intelligence
ACM MobiCom 2022
Pattern-driven compressive sensing for low-power biopotential wearables, enabling efficient on-chip sensing under severe energy and compute constraints.
-
LightLLM: A Versatile Large Language Model for Predictive Light Sensing
ACM SenSys 2025
Fine-tunes LLMs for predictive light sensing by fusing sensor encodings with contextual prompts, while keeping the base model frozen and training only lightweight adapters.
-
FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models
NeurIPS 2025
A cross-domain benchmark for federated fine-tuning of LLMs without sharing raw data, targeting settings where domain-specific and sensitive corpora cannot be centralized.
-
Decoding Ambiguous Emotions with Test-Time Scaling in Audio-Language Models
TASLP 2026
Shows that extra test-time compute improves audio-language models on ambiguous, overlapping affective states that are poorly captured by single-label classification.
-
Test Time Scaling for Auditory Cognition in Audio Language Models
ICASSP 2026
Scales test-time compute to improve auditory cognition in audio language models, such as comprehension and listening recall under noise and overlap.
-
Test Time Adaptation for Speech Emotion Recognition
ICASSP 2026
Adapts speech emotion recognition at inference with unlabeled target data under speaker and cross-corpus shifts, without needing source data or labelled target samples.
-
Disentangling Reasoning in Large Audio-Language Models for Ambiguous Emotion Prediction
Interspeech 2026
Reformulates ambiguous emotion recognition as distributional reasoning in large audio-language models, instead of forcing a single emotion label.