Embedded Arena: Iterative Optimization via Hardware Feedback
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Embedded Arena: Iterative Optimization via Hardware Feedback".
Jane: Here's the gist of the paper. Embedded devices from wildlife monitoring stations to clinical wearables require local AI inference due to latency, communication, or privacy constraints Optimizing models for heterogeneous microcontrollers (MCUs) requires simultaneously satisfying hard physical constraints on memory, power, and temperature while preserving accuracy We ask whether an LLM agent can autonomously navigate this complex, multi-turn pipeline guided by real hardware feedback, and introduce a hardware-in-the-loop agent arena in which the agent iteratively refines both model and firmware—compiling, flashing, and measuring on real hardware—to enable closed-loop optimization Frontier models, including Claude Opus 4.7 and Gemini 3.1 Pro, fail entirely without hardware feedback (0% deployment success), whereas our hardware-in-the-loop formulation achieves the first successful deployment within three iterations and can surpass human expert results within seven This agentic co-optimization achieves 250× compression for vision models with <3.3% accuracy loss and 400× for audio with <6% Feature Error Rate loss, enabling battery-free operation on a commercial MCU via solar harvesting Hardware-in-the-Loop Arena We present a hardware-in-the-loop arena in which an LLM agent must generate, compile, flash, and iteratively refine software subject to physical feedback signals returned by real hardware The formalism we develop in §2.1 is general to any constrained hardware target The agent’s goal is to find a deployable artifact s⋆ that minimizes the task objective: s⋆ = arg min s∈S f(M(s, h)) s.t.
Tom: Φ(s, h) = 1 The benchmark score B for a task is defined as B = f(M(s⋆, h)) Task Instantiations on MCUs The framework is instantiated across three physically distinct axes: model compression, power minimization, and thermal management The agent’s task is to find a deployable artifact s⋆ that minimizes the task objective: s⋆ = arg min s∈S f(M(s, h)) s.t.
Jane: Φ(s, h) = 1 Task 1: Model Compression involves maximizing ondevice inference accuracy macc = Macc(s, D) ∈
0, 1: subject to Φ(s, h) = 1 Task 2: Power Minimization targets the model inference workload with objective fpeak(m) = Ipeak subject to macc ≥ m0 acc Task 3: Thermal Management minimizes peak temperature rise ∆T = Tpeak − Tamb under a controlled ambient load Model Compression Results HIL feedback is necessary for successful deployment, as frontier LLMs fail entirely without it For vision models on the MAX78000, Gemini 3.1 Pro reaches a top mAP50-95 of B = 18.03% The agentic pipeline achieves 250× compression for vision models with <3.3% accuracy loss and 400× for audio with <6% Feature Error Rate loss Power Minimization Analysis Agents may change clock frequencies, sleep scheduling, peripheral gating, and firmware scheduling to minimize peak current or total energy HIL raises successful deployments from 15.9% (Sco) and 38.3% (Doc) to 46.7% (Table 3) for Peak-current minimization on the MAX78000 Thermal Management Evaluation The objective is jointly penalizes peak temperature rise and total task latency: fheat(m) = w∆T ∆T + wτ τ HIL achieves the highest success rates in both conditions, reaching 26.7% at room temperature and 33.3% under contact heating Applications and Real-World Deployments We demonstrate practical impact in two real-world deployments developed in direct collaboration with domain practitioners The HIL pipeline compressed a YOLOv8 elk-detection model to the MAX78000, achieving 96.7% accuracy at under 36 mW during active inference and 7 mW average power Discussion and Conclusion Hardware-in-the-loop feedback raises deployment success from 0% to surpassing human expert results within seven iterations The underlying recipe of reducing a physical objective to a programmatic score that LLM agents can iteratively optimize could generalize well beyond this setting Limitations Our evaluations have three limitations, including the cap of Nmax = 10 iterations due to compounding real resource costs We also evaluated six closed-source frontier models, and our instantiations cover three MCU families and two input modalities Broader Applicability The underlying recipe of reducing a physical objective to a programmatic score that LLM agents can iteratively optimize could generalize well beyond this setting First, additional physical constraints can be folded into the loop, from electromagnetic emissions and vibration tolerance to long-horizon properties like component aging Second, adding a vision frontend to wrap instruments such as mechanical balances, manual titrations, and legacy measurements common in wet labs can enable agents to use a vast array of legacy non-digital measurements Third, this formulation can be extended to scientific domains with unbounded optimization landscapes, such as material design or catalyst discovery References
1: KoelLabs/xlsr-english-01 · Hugging Face, June 2025
3: Andrej.
Tom: karpathy/autoresearch, June 2026
48: John Yang, Carlos E.
Title and authors: Tom: We’re talking about Embedded Arena: Iterative Optimization via Hardware Feedback. The title itself tells you exactly what this paper is about—it’s all about using hardware feedback to iteratively optimize software for embedded devices >
Jane: It suggests a closed loop system where the AI agent isn't just guessing; it's constantly testing and getting real physical feedback on how well its code actually performs on the target hardware >
Lu: The authors are a solid team from places like Washington University and San Diego, all focusing heavily on embedded systems and machine learning representation learning >
Meng: I’m interested in the authors because they seem to have a good grasp of the practical side, since this whole thing is about making things actually deployable in the real world >
Lalam: The paper sets up this arena as a general formalism that works for any constrained hardware target, which is really ambitious because it’s not tied to one specific chip like just one model >
Tom: So what does that mean for us? It means they aren't just solving a problem for the MAXseventy-eight thousand; they are giving us a blueprint that could apply to any MCU family >
Jane: Right, it suggests that the core idea here is translating these complex physical constraints—memory, power, and temperature—into a mathematical score that an AI agent can then optimize toward >
The paper's summary: Tom: The summary points out the core challenge: optimizing models for heterogeneous microcontrollers requires meeting hard physical constraints on memory, power, and temperature simultaneously while keeping accuracy high >
Jane: And the main question they pose is whether an LLM agent can navigate this complex, multi-turn pipeline guided by real hardware feedback to solve that optimization problem autonomously >
Lu: They introduce the hardware in the loop arena where the agent iteratively refines both model and firmware through compiling, flashing, and measuring on real hardware >
Meng: So instead of just optimizing for a fixed number or a benchmark score, they are creating a system where the reward is measured by instruments attached to the device under test >
Lalam: They formalize this as finding an artifact s star that minimizes the task objective subject to satisfying constraints, which they call phi of s and h equaling one >
Tom: That’s a very cool way to frame it; transforming deployment into a constrained optimization problem over real hardware >
The paper's improvements: Jane: One big improvement is that they show that frontier models, like Claude Opus four point seven and Gemini three point one Pro, fail entirely without this hardware feedback loop > <ref:2606.16190#pg1,Claude Opus 4.7 and Gemini 3.1 Pro, fail>
Lu: That’s a strong result because it proves that the current methods used by practitioners to navigate these trade-offs are not robust enough on their own >
Meng: It means that the manual expertise people spend weeks optimizing single models is currently just brittle and hard to scale when you want to deploy things quickly >
Lalam: They show that this agentic co-optimization can achieve two hundred fifty times compression for vision models with less than three point three percent accuracy loss and four hundred times for audio with less than six point zero percent feature error rate loss >
Tom: So the improvements boil down to showing that this method works where previous methods failed, allowing them to find artifacts that actually work on the device, not just those that pass software checks >
Conclusion: Jane: So to wrap up, the main implication is moving away from manual expert optimization toward an AI agent finding a deployable artifact through this closed loop feedback mechanism >
Tom: They’re showing that this system can actually achieve deployment success rates of eighty percent on the hardest tasks compared to zero without hardware feedback >
Lu: It moves the process into a much more accessible workflow where the physical world dictates what we are looking for by making a programmatic score for an artifact that works on hardware >
Meng: From my side, it means we can automate things that used to require weeks of manual tuning, like power minimization and thermal management without needing an expert present >
Lalam: The paper "Embedded Arena: Iterative Optimization via Hardware Feedback" shows how this iterative optimization of models and firmware can lead to real-world impacts like a solar-powered camera trap for elk detection achieving ninety six point seven percent accuracy under seven milliwatts >
Tom: It’s a lot of stuff, but it really shows how we need that kind of system to get complex AI onto the edge where the constraints are this tight >
Zhihan Zhang, Alexander Le Metzger, Jiuyang Lyu, Chun-Cheng Chang, Jiayi Shao, Yujia Liu, Emmanuel Azuh Mensah, Edward Wang Kurtis Heimerl Gregory D. Abowd Shwetak Patel Natasha Jaques Vikram Iyer
University of Washington University of California San Diego Northeastern University
cs.AR, cs.AI
Submitted: 2026-06-15
Updated: 2026-11-06
Comments: NeurIPS 2026 Spotlight: https://ubicomplab.github.io/embedded-arena/
Code: https://github.com/ubicomplab/embedded-arena
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 97/100
The gist: Embedded devices from wildlife monitoring stations to clinical wearables require local AI inference due to latency, communication…
Key concepts
- Hardware-in-the-Loop Arena
- A testing environment where an LLM agent generates code or model changes, compiles them, flashes them onto real microcontrollers (MCUs), and receives physical feedback signals. This allows the agent to close the loop and iteratively optimize software against actual hardware limitations.
- Task Instantiations on MCUs
- Specific optimization goals tailored for resource-constrained devices. These include maximizing inference accuracy, minimizing peak power consumption, and managing thermal rise under a given load. The agent's goal is to find a deployable artifact that minimizes the defined objective function while satisfying all physical constraints.
- Agentic Co-optimization
- Using an LLM agent to autonomously navigate the complex pipeline of model compression, firmware compilation, and hardware testing. The agent treats the deployment process as a search problem, iteratively refining artifacts based on real-world performance metrics rather than relying solely on theoretical or synthetic scores.
Terminology
Summary
Embedded devices from wildlife monitoring stations to clinical wearables require local AI inference due to latency, communication, or privacy constraints Optimizing models for heterogeneous microcontrollers (MCUs) requires simultaneously satisfying hard physical constraints on memory, power, and temperature while preserving accuracy We ask whether an LLM agent can autonomously navigate this complex, multi-turn pipeline guided by real hardware feedback, and introduce a hardware-in-the-loop agent arena in which the agent iteratively refines both model and firmware—compiling, flashing, and measuring on real hardware—to enable closed-loop optimization Frontier models, including Claude Opus 4. p. 0 Optimizing models for heterogeneous microcontrollers (MCUs) requires simultaneously
satisfying hard physical constraints on memory, power, and temperature while
preserving accuracy, a multidimensional optimization that is today performed manually by experts. We ask whether an LLM agent can autonomously navigate this
complex, multi-turn pipeline guided by real hardware feedback, and introduce a
hardware-in-the-loop agent arena in which the agent iteratively refines both model
and firmware—compiling, flashing, and measuring on real hardware—to enable
closed-loop optimization. Frontier models, including Claude Opus 4. 7 and Gemini
3. 1 Pro, fail entirely without hardware feedback (0% deployment success), whereas
our hardware-in-the-loop formulation achieves the first successful deployment
within three iterations and can surpass human expert results within seven. This
agentic co-optimization achieves 250× compression for vision models with <3.Where the paper says it
Where the paper says it
Where the paper says it
Hardware-in-the-Loop Arena
We present a hardware-in-the-loop arena in which an LLM agent must generate, compile, flash, and iteratively refine software subject to physical feedback signals returned by real hardware The formalism we develop in §2.1 is general to any constrained hardware target The agent’s goal is to find a deployable artifact s⋆ that minimizes the task objective: s⋆ = arg min s∈S f(M(s, h)) s. p. 2 , a minimum
accuracy floor). The agent’s goal is to find a deployable artifact s
⋆
that minimizes the task objective:
s
⋆ = arg min
s∈S
f(M(s, h)) s.Where the paper says it
Task Instantiations on MCUs
The framework is instantiated across three physically distinct axes: model compression, power minimization, and thermal management The agent’s task is to find a deployable artifact s⋆ that minimizes the task objective: s⋆ = arg min s∈S f(M(s, h)) s.t. Φ(s, h) = 1
** Task 1: Model Compression involves maximizing ondevice inference accuracy macc = Macc(s, D) ∈ [0, 1] subject to Φ(s, h) = 1 **
** Task 2: Power Minimization targets the model inference workload with objective fpeak(m) = Ipeak subject to macc ≥ m0 acc **
** Task 3: Thermal Management minimizes peak temperature rise ∆T = Tpeak − Tamb under a controlled ambient load **
Model Compression Results
HIL feedback is necessary for successful deployment, as frontier LLMs fail entirely without it For vision models on the MAX78000, Gemini 3.1 Pro reaches a top mAP50-95 of B = 18. p. 5 Table 1
underscores the impact of HIL feedback:
Gemini 3. 1 Pro reaches a top mAP50-95 of
B = 18. 1 Pro, fail entirely without hardware feedback (0% deployment success), whereas
our hardware-in-the-loop formulation achieves the first successful deployment
within three iterations and can surpass human expert results within seven. This
agentic co-optimization achieves 250× compression for vision models with <3.Where the paper says it
Where the paper says it
Power Minimization Analysis
Agents may change clock frequencies, sleep scheduling, peripheral gating, and firmware scheduling to minimize peak current or total energy HIL raises successful deployments from 15.9% (Sco) and 38. p. 4 3 Experiments
To disentangle the contribution of physical feedback from LLMs’ parametric knowledge of embedded
development, we evaluate three MCU targets (ESP32-S3, MAX78000, STM32N6) under three
information settings:
1. Score (Sco): a baseline where the agent receives the task description and initial artifact s0, with
access to a sandboxed environment supporting file manipulation and code execution. At each turn
it observes only the binary feasibility check Φ(st, h) and the scalar optimization objective value
ot = f(M(st, h)). Documentation (Doc): the agent is additionally provided expert-curated device datasheets and
toolchain documentation describing target constraints (e.Where the paper says it
Where the paper says it
Thermal Management Evaluation
The objective is jointly penalizes peak temperature rise and total task latency: fheat(m) = w∆T ∆T + wτ τ HIL achieves the highest success rates in both conditions, reaching 26.7% at room temperature and 33. p. 4 The agent minimizes peak temperature rise ∆T = Tpeak − Tamb under a
controlled ambient load, with Tmax as a hard safety ceiling in h. Because ∆T depends on Tamb,
we evaluate each device across multiple controlled ambient conditions, from room temperature to a
surface approximating skin contact; details are in Appendix E.Where the paper says it
Applications and Real-World Deployments
We demonstrate practical impact in two real-world deployments developed in direct collaboration with domain practitioners The HIL pipeline compressed a YOLOv8 elk-detection model to the MAX78000, achieving 96.7% accuracy at under 36 mW during active inference and 7 mW average power
Discussion and Conclusion
Hardware-in-the-loop feedback raises deployment success from 0% to surpassing human expert results within seven iterations The underlying recipe of reducing a physical objective to a programmatic score that LLM agents can iteratively optimize could generalize well beyond this setting
Limitations
Our evaluations have three limitations, including the cap of Nmax = 10 iterations due to compounding real resource costs We also evaluated six closed-source frontier models, and our instantiations cover three MCU families and two input modalities
Broader Applicability
The underlying recipe of reducing a physical objective to a programmatic score that LLM agents can iteratively optimize could generalize well beyond this setting First, additional physical constraints can be folded into the loop, from electromagnetic emissions and vibration tolerance to long-horizon properties like component aging Second, adding a vision frontend to wrap instruments such as mechanical balances, manual titrations, and legacy measurements common in wet labs can enable agents to use a vast array of legacy non-digital measurements Third, this formulation can be extended to scientific domains with unbounded optimization landscapes, such as material design or catalyst discovery
References
[1] KoelLabs/xlsr-english-01 · Hugging Face, June 2025
[3] Andrej. karpathy/autoresearch, June 2026
[48] John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press
[52] Barret Zoph and Quoc V. Le. Neural Architecture Search with Reinforcement Learning
[47] Huanqi Yang, Mingzhe Li, Mingda Han, Zhenjiang Li, and Weitao Xu
[51] Zhihan Zhang, Alexander Metzger, Yuxuan Mei, Felix Hähnlein, Zachary Englhardt
[49] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu
[33] Ji Lin, Wei-Ming Chen, Yujun Lin, John Cohn, Chuang Gan, and Song Han
[45] Xiaoye Xie, Xinzhu Liu, Jianhua Cai, Bohan Zhang, Tianjun Sun
[38] A. R. Moritz and F. C. Henriques
[44] Anran Wang, Vikram Iyer, Vamsi Talla, Joshua R. Smith
[35] Mike A. Merrill, Alexander Glenn Shaw, Nicholas Carlini
[47] Huanqi Yang, Mingzhe Li, Mingda Han, Zhenjiang Li, and Weitao Xu
[48] John Yang, Carlos E. Jimenez, Alexander Wettig
[53] Barret Zoph and Quoc V. Le
Appendix Table of Contents
A Background 15
A.1 MCU Deployment Regime. p. 4 , supported operators, memory limits,
required configuration formats) as part of s0 and, except for model compression, may use web
search to look up further details. This baseline is motivated by our finding that without such
materials, frontier LLMs fail even basic MCU deployment questions (Appendix B).Where the paper says it
B Benchmarking LLM Knowledge of MCU Deployment 17
C Model Compression 18
C.3 Failure Modes in Model Compression. p. 7 We observe agents
disabling cameras, substituting dummy inputs, reducing UDP fan-out, shortening prompts, and
hard-coding completion messages. HIL feedback addresses these failure modes by exposing hidden execution state through physical
traces, serial logs, checkpoint detection, and behavior judging, enabling agents to first establish a
valid hardware baseline and then optimize rather than chasing artifacts that only work in software.Where the paper says it
D Power Minimization 22
E Thermal Management 26
F Structured Natural Language Feedback Example 30
G MAX78000 MCU 32
H Wildlife Monitoring: Elk Detection 34
I Early Language Development and Phoneme Perception 39
Appendix A Background
A.1 MCU Deployment Regime MCUs occupy a distinct position in the AI deployment spectrum with orders-of-magnitude differences in both compute capabilities, energy consumption and cost Compared even to edge platforms like the Raspberry Pi 5, MCUs offer three to five orders of magnitude less memory while consuming roughly 500× less power The compute constraints however require aggressive compression of popular pre-trained models which exceed typical MCU memory budgets by 200 to 3,000×
Appendix A.2 Related Work
TinyML and MCU Deployment Machine learning on resource-constrained microcontrollers has grown rapidly over the past decade, driven by the proliferation of edge devices and the need for on-device intelligence Dominant approaches include manually designed compact architectures [41], Neural Architecture Search (NAS) [52, 42, 7, 24, 33, 5], and post-training compression via quantization, pruning, and knowledge distillation [21, 37, 32, 23] Automated deployment frameworks such as LiteRT (formerly TensorFlow Lite Micro) [13], microTVM [10], and STM32Cube.AI translate trained models into MCU-targeted firmware, handling operator scheduling, memory planning, and quantized kernel selection These approaches share two limitations that motivate our work: First, they optimize within the tensor arena and treat memory as the sole constraint Second, they cast deployment as a closed numerical optimization, with a fixed search space (architectures, bitwidths, sparsity patterns) and dense gradient or score signals
Appendix B Benchmarking LLM Knowledge of MCU Deployment
All evaluated models struggle on this benchmark Even with web search and context provided, only GPT-5 surpasses 60% accuracy (Table 5)
Appendix C Model Compression
Manual inspection of compression trajectories shows that the main bottleneck is deployability rather than model construction Agents frequently produce neural networks that are syntactically valid and train to nontrivial validation scores, but fail when translated through the target-specific synthesis and deployment stack
Appendix D Power Minimization
Agent-driven power optimization frequently fails due to the profound disconnect between syntactic software correctness and complex hardware state management We identified six primary failure modes, including Reward Hacking (Workload Evasion) and Execution State Blindness
Appendix E Thermal Management
Because thermal dissipation is intrinsically linked to the underlying power profile (Q = R τ 0 P(t) dt), the thermal optimization task on the ESP32-S3R8 shares the fundamental failure modes detailed in Power Minimization
Appendix F Structured Natural Language Feedback Example
The feedback is fully automated and does not depend on human intervention Below is an example observation from a MAX78000 peak-current minimization run
Appendix G MAX78000 MCU
Contrary to intuition, peak current consumption shows minimal correlation with model size Total inference time and power consumption scale approximately linearly with model size
Appendix H Wildlife Monitoring: Elk Detection
Our HIL pipeline compressed a YOLOv8 elk-detection model to the MAX78000, achieving 96.7% accuracy at under 36 mW during active inference and 7 mW average power The HIL pipeline jointly optimized compression, power, and thermal on the STM32N6, compressing a Wav2Vec2-based speech-to-IPA model to a Feature Error Rate 8. p. 8 Children’s wearables impose two hard physical constraints beyond accuracy: the device must run all
day on a tiny 1,200 mAh battery, and surface temperatures must remain within safe contact limits
throughout. Our HIL pipeline jointly optimized compression, power, and thermal on the STM32N6,
compressing a Wav2Vec2-based speech-to-IPA model to a Feature Error Rate 8.Where the paper says it
Appendix I Early Language Development and Phoneme Perception
Our HIL pipeline jointly optimized compression, power, and thermal on the STM32N6, compressing a Wav2Vec2-based speech-to-IPA model to a Feature Error Rate 8.44%, while jointly satisfying both constraints
MCUs provide 3-5 orders of magnitude less memory and consume 500× less power even compared to common edge devices like the Raspberry Pi 5 However, this efficiency comes at a cost: popular pre-trained models exceed the memory budget by 200-3,000×
LLM agents are well-suited to this regime, where iterative exploration benefits from semantic priors over code and hardware behavior rather than gradient signals over a fixed parameterization Prior work exploring and benchmarking LLMs for embedded systems focuses on single-turn or software-only iterations, with a focus on generic tasks such as driver generation and system integration
Agents frequently produce neural networks that are syntactically valid and train to nontrivial validation scores, but fail when translated through the target-specific synthesis and deployment stack
**Appendix A.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Hardware-in-the-loop agent arena enables closed-loop optimization by allowing agents to
iteratively refine both model and firmware—compiling, flashing, and measuring on real hardware.
This achieves a deployment success rate of 80% on the hardest tasks compared to 0% without hardware feedback. -
The system can achieve significant model compression, such as "250x compression for vision models with <3.3% accuracy loss," while simultaneously reducing power draw by over 58%.
-
Agents can autonomously navigate complex deployment pipelines by minimizing a task objective, formalized as
s⋆ = arg min s∈S f(M(s, h)) s.t. Φ(s, h) = 1,
transforming the process into anaccessible workflow.
-
The system can perform power minimization by optimizing objectives such as
peak current minimization targets the model inference workload
andenergy minimization targets a fixed-payload wireless transmission.
-
The system can manage thermal constraints by minimizing a joint objective, as defined by
fheat(m) = w∆T ∆T + wτ τ,
which penalizes both peak temperature rise and total task latency. -
The improved system can detect and correct hardware-specific failure modes, such as
Cross-Ecosystem Hallucination
andReward Hacking (Workload Evasion),
by using physical traces and serial logs to bridge the gap between syntactic correctness and actual hardware execution. -
The agent can overcome premature convergence by being guided by physical feedback, which allows it to move beyond baseline code that only passes software checks to find artifacts that
actually work on the device.
-
The system can be applied to real-world systems, such as a
solar-powered camera trap for elk detection achieving 96.7% accuracy under 7 mW average power
or achildren’s head-worn device with 8.44% FER for phonetic transcription.
Abstract
As AI moves into the physical world, from robots to environmental sensors to medical wearables, it increasingly relies on local inference to meet latency, communication, or privacy constraints. Optimizing models for heterogeneous edge hardware such as microcontrollers (MCUs) requires simultaneously satisfying hard physical constraints on memory, power, and temperature while preserving accuracy, a multidimensional optimization that is today performed manually by experts. We ask whether an LLM agent can autonomously navigate this complex, multi-turn pipeline guided by real hardware feedback, and introduce a hardware-in-the-loop agent arena in which the agent iteratively refines both model and firmware by compiling, flashing, and measuring on real hardware, enabling closed-loop optimization. Frontier models, including Claude Opus 4.7 and Gemini 3.1 Pro, fail entirely without hardware feedback (0% deployment success), whereas our hardware-in-the-loop formulation achieves the first successful deployment within three iterations and can surpass human expert results within seven. This agentic co-optimization achieves 250x compression for vision models with <3.3% accuracy loss and 400x for audio with <6% Feature Error Rate loss, enabling battery-free operation on a commercial MCU via solar harvesting. We demonstrate practical impact in two real-world systems: an elk-detection camera trap (96.7% accuracy) and a phonetic-transcription wearable (8.37% FER) for child development research.
Sources
- MicroNets: Neural Network Architectures for Deploying TinyML Applications on Commodity Microcontrollers
- ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware
- Evaluating Large Language Models Trained on Code
- Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering Tasks with Generative Optimization
- TensorFlow Lite Micro: Embedded Machine Learning on TinyML Systems
- Distilling the Knowledge in a Neural Network
- Towards a Science of Scaling Agent Systems
- Skilled AI Agents for Embedded and IoT Systems Development
- MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement
- EmbedAgent: Benchmarking Large Language Models in Embedded System Development
- Neural Architecture Search with Reinforcement Learning
Related papers
- WitCert: Sound Runtime Risk Observability and Gating for KV-Cache Quantization
- SparsePixels: Efficient Convolution for Sparse Data on FPGAs
- Golden Ruler: A Numeric Format Catalog with Bit-Exact Conformance Vectors for FP8, BF16, MXFP4, and Microscaling Formats
- AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving
- RFChipAgent: Multi-Agentic AI Flow for Analog/RF Chip Design
- PoisonCap: Efficient Hierarchical Temporal Safety for CHERI