FairCompressAgent: An Agentic Framework for Fairness-Aware Model Compression for FPGA Deployment

# FairCompressAgent: An Agentic Framework for Fairness-Aware Model Compression for FPGA Deployment
Source: [https://arxiv.org/html/2609.17786](https://arxiv.org/html/2609.17786)
Yuanbo GuoAffiliation:Department of Computer Science and Engineering University of Notre Dame Notre Dame, USA yguo6@nd\.eduYiyu ShiAffiliation:Department of Computer Science and Engineering University of Notre Dame Notre Dame, USA yshi4@nd\.edu

###### Abstract

Fairness\-aware model compression requires selecting methods and configurations that balance accuracy, fairness, and deployment cost\. These decisions become more difficult when compression methods are composed or the user’s requirements change\. In this paper, we propose FairCompressAgent \(FCA\), an agentic framework that integrates fairness\-aware pruning, incremental quantization, and sparse low\-rank factorization through a common operator interface\. A language\-model planner uses model profiles and measured outcomes to select compression configurations, while an execution layer performs compression, fine\-tuning, evaluation, and constraint\-based selection\. FCA also supports requirement updates and reports the remaining violation when a request cannot be satisfied\. Experiments on Fitzpatrick\-17k with VGG\-11 compare four search methods over 40 measured configurations\. Under the accuracy\-constrained request, FCA selects a compressed model with 59\.54% less inference tensor storage, while validation average precision increases from 0\.5141 to 0\.5233 and equalized opportunity \(EOpp\) decreases from 0\.2251 to 0\.2168\. It reaches the same final selection as one\-shot planning with 7\.33 versus 12 candidate evaluations on average, under their respective stopping policies\. Repeated fine\-tuning, held\-out testing, and online requirement updates characterize the stability and interactive use of this compression workflow\. The results demonstrate how measured feedback and explicit constraints support the selection and interactive refinement of fairness\-aware compression configurations\.

###### Index Terms:

deep learning, fairness, hardware efficiency, AI agent, FPGA\.

## IIntroduction

As deep learning \(DL\) becomes increasingly adopted in different areas of daily life, its applications raise a variety of issues that need to be addressed\. Among these concerns, fairness and efficiency are often treated as separate problems\. To begin with, fairness concerns arise from discrimination in prediction outcomes, such as employment discrimination\[[1](https://arxiv.org/html/2609.17786#bib.bib10),[2](https://arxiv.org/html/2609.17786#bib.bib11)\], as well as disparities in prediction quality, such as healthcare inequity\[[2](https://arxiv.org/html/2609.17786#bib.bib11),[3](https://arxiv.org/html/2609.17786#bib.bib12)\]\. DL models can influence people’s decisions, so applications supported by unfair models may contribute to systematic discrimination and disadvantage particular demographic groups\. To deal with fairness issues, previous works majorly focus on algorithm optimization\[[4](https://arxiv.org/html/2609.17786#bib.bib29),[5](https://arxiv.org/html/2609.17786#bib.bib30),[6](https://arxiv.org/html/2609.17786#bib.bib34)\]\. Meanwhile, model compression is one of the most useful approaches for deploying DL models locally with limited hardware resources\[[7](https://arxiv.org/html/2609.17786#bib.bib23)\], including FPGA platforms\. For instance, pruning\[[8](https://arxiv.org/html/2609.17786#bib.bib13),[9](https://arxiv.org/html/2609.17786#bib.bib14)\]and quantization\[[10](https://arxiv.org/html/2609.17786#bib.bib15),[11](https://arxiv.org/html/2609.17786#bib.bib16)\]are two widely used techniques for model compression\. Conventionally, however, compression decisions mainly balance prediction quality and computational cost, without explicitly considering unequal performance across demographic groups\.\[[12](https://arxiv.org/html/2609.17786#bib.bib24)\]also pointed out that model compression could disproportionately affect bias even when accuracy performance changes little\.

On top of that, FPGA deployment is particularly relevant to model compression\. Available compute resources and memory bandwidth jointly constrain accelerator design, while on\-chip storage influences data reuse and transfers to external memory\[[13](https://arxiv.org/html/2609.17786#bib.bib21)\]\. Low\-precision designs further connect numerical representation with parallelism and dataflow\[[14](https://arxiv.org/html/2609.17786#bib.bib22)\]\. Different compression methods address these constraints in different ways\. For example, pruning changes connectivity of neurons, while quantization restricts how weights are represented\. The resulting representation determines which storage and execution opportunities a target design can exploit\. Choosing a suitable compression configuration therefore requires considering operator compatibility and resource requirements together with inference performance in terms of both accuracy and fairness\[[15](https://arxiv.org/html/2609.17786#bib.bib28)\]\.

Recent works start to explore the relationship between DL model fairness and hardware platforms\[[16](https://arxiv.org/html/2609.17786#bib.bib31),[17](https://arxiv.org/html/2609.17786#bib.bib32),[18](https://arxiv.org/html/2609.17786#bib.bib35),[19](https://arxiv.org/html/2609.17786#bib.bib36)\]\. Specifically, fairness\-aware compression methods have been introduced to seek fairness improvements during compression, creating a trade\-off among accuracy, fairness, and efficiency\. FairPrune\[[20](https://arxiv.org/html/2609.17786#bib.bib1)\]addresses this problem through group\-sensitive pruning, and FairQuantize\[[21](https://arxiv.org/html/2609.17786#bib.bib2)\]extends the idea to incremental weight quantization\. In addition, sparse low\-rank factorization offers another way to reduce model complexity\[[22](https://arxiv.org/html/2609.17786#bib.bib3)\], and FairLRF adapts it to group\-sensitive compression\[[23](https://arxiv.org/html/2609.17786#bib.bib17)\]\. There have also been a few works following these paths\[[24](https://arxiv.org/html/2609.17786#bib.bib33)\]\. While all these methods provide different ways to address fairness through compression, choosing among them still requires decisions about compression strength, fine\-tuning, layer coverage, and method order\.

Joint compression search and language\-model\-assisted optimization provide foundations for coordinating these decisions\. APQ jointly searches architecture, pruning, and quantization policies\[[25](https://arxiv.org/html/2609.17786#bib.bib6)\], while LLAMBO investigates language models within Bayesian optimization\[[26](https://arxiv.org/html/2609.17786#bib.bib7)\]\. FairAgent uses language\-model agents to support fairness\-aware machine learning\[[27](https://arxiv.org/html/2609.17786#bib.bib8)\], and ProfilingAgent uses profiling results to guide model optimization\[[28](https://arxiv.org/html/2609.17786#bib.bib9)\]\. Applying these ideas to fairness\-aware compression requires a common interface for compatible operators and explicit criteria for evaluating accuracy, fairness, and efficiency\.

To integrate these capabilities into comprehensive workflows, we propose FairCompressAgent \(FCA\), an agentic framework for selecting and composing fairness\-aware compression operators under user\-defined requirements\. FCA uses model profiles and evaluation results to guide compression choices and allows the user to revise requirements during the search\. Figure[1](https://arxiv.org/html/2609.17786#S1.F1)summarizes the interaction between the user and FCA\.

Specify a model, annotated data, and requirementsAccuracy, fairness, storage, and search budgetFCA evaluates compression configurationsCompare measured trade\-offs under the requirementsInspect the selected model and reportReview performance, resource use, and unmet constraintsAccept the result or revise the requirementsContinue with the accumulated observationsFig\. 1:User interaction with FCA\. The user supplies the task and requirements, inspects measured results, and can revise the bounds to continue with the accumulated evaluation records\.The main contributions of this paper are as follows:

- •We formulate fairness\-aware compression as constrained configuration selection and organize pruning, quantization, and sparse low\-rank factorization through a common perturbation view and executable interface\.
- •We develop an agent architecture for adaptive compression planning, with explicit constraint checks and support for requirement updates\.
- •We evaluate the framework using 40 measured configurations, four search methods, repeated fine\-tuning, held\-out testing, and fresh online execution\. FCA selects a model with 59\.54% less inference tensor storage under the accuracy\-constrained request and identifies the unmet constraint under a stricter fairness request\.

The source code for FairCompressAgent is available at[https://github\.com/guoyb17/FairCompressAgent](https://github.com/guoyb17/FairCompressAgent)\.

## IIProblem Formulation

### II\-AInputs and Accuracy Metrics

Consider a classification dataset𝒟=\{\(xi,yi,ai\)\}i=1N\\mathcal\{D\}=\\\{\(x\_\{i\},y\_\{i\},a\_\{i\}\)\\\}\_\{i=1\}^\{N\}, wherexix\_\{i\}is the input,yiy\_\{i\}is the class label, andaia\_\{i\}is a sensitive attribute\. We focus on binary sensitive attributes, which divide the annotated samples into groupsg∈\{0,1\}g\\in\\\{0,1\\\}\. Given a pre\-trained modelf⁡\(θ0\)f\(\\theta\_\{0\}\), our objective is to find a compressed model that satisfies the user’s requirements on accuracy, fairness, and efficiency\.

Following FairPrune, FairQuantize and FairLRF, we use average precision as the main measure of accuracy performance\. It is the mean of the macro precision values of the two demographic groups\. For a fixed set ofKKclasses, it is defined as

Pg=1K​∑k=1KPrecg,k,Pavg=P0\+P12,P\_\{g\}=\\frac\{1\}\{K\}\\sum\_\{k=1\}^\{K\}\\operatorname\{Prec\}\_\{g,k\},\\qquad P\_\{\\mathrm\{avg\}\}=\\frac\{P\_\{0\}\+P\_\{1\}\}\{2\},\(1\)wherePrecg,k\\operatorname\{Prec\}\_\{g,k\}is the one\-versus\-rest precision of classkkin groupgg\. Precision is set to zero when no sample in that group is predicted as classkk\.

### II\-BFairness and Efficiency Metrics

Equalized opportunity and equalized odds\[[29](https://arxiv.org/html/2609.17786#bib.bib4)\]have been widely used in previous research regarding fairness\. For equalized opportunity, it is the difference of true negative rates \(denoted as EOpp0\) or true positive rates \(denoted as EOpp1\) of different demographic groups within each class of target attributes, followed by sum of them\. For equalized odds \(denoted as EOdd\), it collects the difference of true positive rates plus difference of false positive rates\. These metrics can be calculated by the following equations:

E​O​p​p​0=∑k=1K\|T​N​Rk1−T​N​Rk0\|,EOpp0=\\sum\_\{k=1\}^\{K\}\{\\lvert TNR\_\{k\}^\{1\}\-TNR\_\{k\}^\{0\}\\rvert\},\(2\)E​O​p​p​1=∑k=1K\|T​P​Rk1−T​P​Rk0\|,EOpp1=\\sum\_\{k=1\}^\{K\}\{\\lvert TPR\_\{k\}^\{1\}\-TPR\_\{k\}^\{0\}\\rvert\},\(3\)E​O​d​d=∑k=1K\|T​P​Rk1−T​P​Rk0\+F​P​Rk1−F​P​Rk0\|\.EOdd=\\sum\_\{k=1\}^\{K\}\{\\lvert TPR\_\{k\}^\{1\}\-TPR\_\{k\}^\{0\}\+FPR\_\{k\}^\{1\}\-FPR\_\{k\}^\{0\}\\rvert\}\.\(4\)For all of EOpp0, EOpp1, and EOdd, the smaller the better\.

These metrics reflect the statistical differences of model performance on different demographic groups from different perspectives\. Given that fairness for positive cases is usually more important in real world scenarios like disease diagnosis, we use EOpp1 as the representation of fairness metrics in this paper, denoted as EOpp below\.

A compressed model is represented byz=\(θ,ρ\)z=\(\\theta,\\rho\), whereρ\\rhospecifies its executable form, including factor layers and constrained weights\. We measure storageS⁡\(ρ\)S\(\\rho\)as the bytes occupied by inference parameter and buffer tensors, excluding optimizer state, training masks, and file headers\. The storage measurement therefore reflects the representation used for inference\. Execution cost additionally depends on how the representation maps to the target architecture\. The present experiments use dense FP32 tensors, including the factors of low\-rank layers\.

### II\-CConstrained Configuration Selection

A configurationπ=\(\(o1,η1\),…,\(oT,ηT\)\)\\pi=\(\(o\_\{1\},\\eta\_\{1\}\),\\ldots,\(o\_\{T\},\\eta\_\{T\}\)\)specifies compression operatorsoto\_\{t\}, their order, and settingsηt\\eta\_\{t\}\. Each configuration starts from the supplied baseline and produces a candidatezπz\_\{\\pi\}\. The three objectives are higherPavgP\_\{\\mathrm\{avg\}\}, lower EOpp, and smaller storage:

minπ⁡\(−Pavg​\(zπ\),EOpp⁡\(zπ\),S⁡\(ρπ\)\)\.\\min\_\{\\pi\}\\bigl\(\-P\_\{\\mathrm\{avg\}\}\(z\_\{\\pi\}\),\\ \\mathrm\{EOpp\}\(z\_\{\\pi\}\),\\ S\(\\rho\_\{\\pi\}\)\\bigr\)\.\(5\)The user specifies acceptable average precision lossϵP\\epsilon\_\{P\}, an EOpp upper bounddmaxd\_\{\\max\}, and, optionally, a storage\-ratio upper boundsmaxs\_\{\\max\}:

Pavg​\(zπ\)\\displaystyle P\_\{\\mathrm\{avg\}\}\(z\_\{\\pi\}\)≥Pavg​\(z0\)−ϵP,\\displaystyle\\geq P\_\{\\mathrm\{avg\}\}\(z\_\{0\}\)\-\\epsilon\_\{P\},\(6\)EOpp⁡\(zπ\)\\displaystyle\\mathrm\{EOpp\}\(z\_\{\\pi\}\)≤dmax,\\displaystyle\\leq d\_\{\\max\},\(7\)S⁡\(ρπ\)/S⁡\(ρ0\)\\displaystyle S\(\\rho\_\{\\pi\}\)/S\(\\rho\_\{0\}\)≤smax\.\\displaystyle\\leq s\_\{\\max\}\.\(8\)Among feasible observations, the experiments select minimum storage, then maximumPavgP\_\{\\mathrm\{avg\}\}, then minimum EOpp, with candidate identifiers breaking any remaining ties\.

## IIIFairness\-Aware Compression Operators

### III\-AA Common Group\-Sensitive Score

Different weights can contribute unequally to the prediction performance of different demographic groups, and second\-order sensitivity provides a promising basis for ranking weight perturbations\[[30](https://arxiv.org/html/2609.17786#bib.bib25)\]\. FairPrune and FairQuantize use this observation to decide which weights to modify, while FairLRF applies a related score to low\-rank factors\[[20](https://arxiv.org/html/2609.17786#bib.bib1),[21](https://arxiv.org/html/2609.17786#bib.bib2),[23](https://arxiv.org/html/2609.17786#bib.bib17)\]\. We organize these existing methods as compression operators with a common scoring and evaluation interface\.

LetLg​\(θ\)L\_\{g\}\(\\theta\)be the mean loss on a fixed scoring subset for groupgg\. For a weight perturbationΔ\\Delta, the Taylor expansion gives

Lg\(θ\+Δ\)−Lg\(θ\)=∇Lg⊤Δ\+12Δ⊤HgΔ\+Rg\(Δ\),L\_\{g\}\(\\theta\+\\Delta\)\-L\_\{g\}\(\\theta\)=\\nabla L\_\{g\}^\{\\top\}\\Delta\+\\tfrac\{1\}\{2\}\\Delta^\{\\top\}H\_\{g\}\\Delta\+R\_\{g\}\(\\Delta\),\(9\)whereHgH\_\{g\}is the group\-specific Hessian andRgR\_\{g\}contains higher\-order terms\. The operators use the following diagonal quadratic score:

si=12​Δi2​\(hu,i−β​hp,i\),β≥0,s\_\{i\}=\\tfrac\{1\}\{2\}\\Delta\_\{i\}^\{2\}\\left\(h\_\{u,i\}\-\\beta h\_\{p,i\}\\right\),\\qquad\\beta\\geq 0,\(10\)Herehg,ih\_\{g,i\}is theiith diagonal entry ofHgH\_\{g\}, anduuandppare the reference underperforming and better\-performing groups\. The experiment fixes these roles as Light and Dark from the baseline comparison throughout search\. A smaller score favors changes with less estimated loss impact on groupuurelative to grouppp\. The hyper\-parameterβ\\betaadjusts the local ranking, while \([7](https://arxiv.org/html/2609.17786#S2.E7)\) governs the measured fairness of the resulting model\.

The local score ranks proposed modifications\. Full evaluation after fine\-tuning determines whether the resulting model satisfies the user’s requirements\.

### III\-BPruning and Incremental Quantization

For FairPrune, removing a weight sets it to zero, givingΔi=−θi\\Delta\_\{i\}=\-\\theta\_\{i\}in \([10](https://arxiv.org/html/2609.17786#S3.E10)\)\[[20](https://arxiv.org/html/2609.17786#bib.bib1)\]\. The operator selects the required number of lowest\-scoring eligible weights within each target tensor\. A cumulative mask keeps these weights at zero during subsequent fine\-tuning\.

For FairQuantize, the displacement isΔi=Q⁡\(θi\)−θi\\Delta\_\{i\}=Q\(\\theta\_\{i\}\)\-\\theta\_\{i\}\[[21](https://arxiv.org/html/2609.17786#bib.bib2)\]\. The implementation uses a signed power\-of\-two mapping,

Q⁡\(w\)=\{0,w=0,sign⁡\(w\)​2round⁡\(log2⁡\|w\|\),w≠0\.Q\(w\)=\\begin\{cases\}0,&w=0,\\\\ \\operatorname\{sign\}\(w\)\\,2^\{\\operatorname\{round\}\(\\log\_\{2\}\|w\|\)\},&w\\neq 0\.\\end\{cases\}\(11\)The squared displacement in \([10](https://arxiv.org/html/2609.17786#S3.E10)\) accounts for the size of each quantization change\. Following incremental network quantization\[[31](https://arxiv.org/html/2609.17786#bib.bib5)\], the selected values are frozen while the remaining weights receive fine\-tuning updates\. Previously frozen entries remain fixed during subsequent increments\. The resulting power\-of\-two weights provide a representation that can be mapped to a compatible quantized execution backend\.

### III\-CSparse Low\-Rank Factorization

For a linear layer withmminputs andnnoutputs, write its weight matrix in input\-by\-output coordinates asW∈ℝm×nW\\in\\mathbb\{R\}^\{m\\times n\}\. Truncated singular value decomposition gives

W≈Ur​Σr​Vr⊤,W\\approx U\_\{r\}\\Sigma\_\{r\}V\_\{r\}^\{\\top\},\(12\)whererris the retained rank\. Sparse low\-rank factorization further removes selected factor entries\[[22](https://arxiv.org/html/2609.17786#bib.bib3)\]\. FairLRF\[[23](https://arxiv.org/html/2609.17786#bib.bib17)\]determines where this sparsity is applied using group\-sensitive scores\. Letq=⌊α​r⌋q=\\lfloor\\alpha r\\rfloorbe the retained prefix of singular directions\. For rowiiofUrU\_\{r\}, its tail\-removal score is

SiU=∑j=q\+1r12​Ui​j2​\(hu,i​jU−β​hp,i​jU\)\.S\_\{i\}^\{U\}=\\sum\_\{j=q\+1\}^\{r\}\\tfrac\{1\}\{2\}U\_\{ij\}^\{2\}\\left\(h\_\{u,ij\}^\{U\}\-\\beta h\_\{p,ij\}^\{U\}\\right\)\.\(13\)The corresponding expression scores the tail of each column ofVr⊤V\_\{r\}^\{\\top\}\. The lowest\-scoring rows and columns receive tail removal according to a sparsification fraction\. Curvature is computed in the factor coordinates with the singular values fixed during scoring\. We use the quadratic tail sum in \([13](https://arxiv.org/html/2609.17786#S3.E13)\) throughout this study\.

After sparsification, singular values are absorbed into one factor, and inference uses two consecutive linear maps with the original bias applied at the output\. Their matrix storage isr⁡\(m\+n\)r\(m\+n\)FP32 elements, compared withm​nmnfor the original layer\. The factor masks are preserved during fine\-tuning\. The evaluated factorization operator targets fully connected layers\.

### III\-DComposition of Operators

We denote the three operators byPP,QQ, andLL\. Besides individual operators, FCA supportsP→QP\\\!\\rightarrow\\\!QandL→QL\\\!\\rightarrow\\\!Q\. ForP→QP\\\!\\rightarrow\\\!Q, quantization acts on the remaining unpruned weights, and all pruning zeros are preserved\. ForL→QL\\\!\\rightarrow\\\!Q, the quantizer operates directly on the executable factors and recomputes curvature after the low\-rank transformation and its fine\-tuning stage\. Factor sparsity remains fixed during quantization and further fine\-tuning\. The executor restores frozen values after each optimizer step and clears their gradients and momentum updates\.

Order matters because the same representation property need not survive the reverse composition\. For example, factorizing a quantized matrix generally produces factors outside its original quantization codebook\. Similarly, factorizing a pruned matrix need not preserve its zeros\. These reverse orders are therefore excluded from the present configuration catalogue\.

### III\-EOperator Information for Planning

The planner receives each operator’s supported layers, settings, and representation changes\. For a linear layer, two dense factors reduce matrix storage whenr<m​n/\(m\+n\)r<mn/\(m\+n\)\. The executor exposes storage for the retained rank, while evaluation supplies average precision and EOpp after fine\-tuning\. Pruning masks and quantized values specify constraints for subsequent transformations\. This information lets FCA compare individual methods and compatible compositions under the same selection rule\.

## IVFCA Architecture

Fig\. 2:Overall architecture of FCA\. The planner selects configurations using the model profile, user requirements, and accumulated evaluation records\. The operator library supplies compatible transformations to the executor, and the evaluator returns measured outcomes to the shared session record\. Constraint\-based selection produces the chosen model and report; the feedback path supports subsequent planning and requirement updates\.### IV\-APlanning, Execution, and Evaluation

Reasoning\-and\-action agents provide a precedent for interleaving planning with tool feedback\[[32](https://arxiv.org/html/2609.17786#bib.bib26)\]\. Figure[2](https://arxiv.org/html/2609.17786#S4.F2)presents the three main components of FCA: a language\-model planner, a compression executor, and an evaluator\. The planner receives baseline performance, group coverage, layer dimensions, supported configurations, and the user’s constraints and search budget\. It selects from a finite catalog of individual operators and compatible compositions\. Each configuration specifies the operator sequence, target layers, compression strength, and group\-weighting parameterβ\\beta\. The planner combines structural information, such as the relationship between retained rank and tensor storage, with queried evaluation results to select a configuration and explain its choice\.

The executor handles model transformations and fine\-tuning, while the evaluator measuresPavgP\_\{\\mathrm\{avg\}\}, EOpp, group coverage, inference tensor storage, and execution cost\. Typed tool interfaces restrict the planner to supported configurations and operations, including evaluation, comparison, requirement updates, and report export\. Images, individual predictions, and model weights remain local; compact profiles and queried outcomes enter the planner’s context\.

### IV\-BState and Decision Cycle

The session state at stepttconsists of the current requirements𝒞t\\mathcal\{C\}\_\{t\}, remaining budgetbtb\_\{t\}, and evaluation recordsEtE\_\{t\}\. Each record stores the configuration, measured metrics, execution cost, completion status, and saved model reference\.

Before an evaluation, FCA checks that the requested configuration is supported and that sufficient budget remains\. A repeated request retrieves its existing record\. For a new configuration in online execution, the executor performs compression, fine\-tuning, and validation from the supplied baseline\. FCA then updatesEtE\_\{t\}andbtb\_\{t\}and applies the selection rule in Section[II](https://arxiv.org/html/2609.17786#S2)under𝒞t\\mathcal\{C\}\_\{t\}\. Failed evaluations remain in the record with their status\.

### IV\-CMeasured Feedback and Requirement Updates

For controlled comparison of search methods, FCA retrieves measured outcomes and execution times from a shared pre\-evaluated candidate set\. The planner receives only queried outcomes, and each distinct query contributes its recorded execution time to the accumulated model evaluation time\. This replay protocol keeps the candidate measurements identical across search methods\.

When the user revises a requirement, the update tool records the instruction, changes𝒞t\\mathcal\{C\}\_\{t\}, and reapplies the selection rule toEtE\_\{t\}without further model execution\. Each reported selection retains the requirements under which it was made\.

### IV\-DStopping and Unmet Constraints

Each session has limits on distinct evaluations, accumulated model evaluation time, and language\-model requests\. The planner may stop earlier when the observed trade\-offs are sufficient for its response\.

When the observed set contains no feasible candidate, FCA reports the unmet request and identifies the smallest normalized violation for diagnosis:

v⁡\(z\)=\[Pavg​\(z0\)−Pavg​\(z\)−ϵPϵP\]\+\+\[EOpp⁡\(z\)−dmaxdmax\]\+\+𝟏S​\[S⁡\(ρ\)/S⁡\(ρ0\)−smaxsmax\]\+,\\begin\{split\}v\(z\)=\{\}&\\left\[\\frac\{P\_\{\\mathrm\{avg\}\}\(z\_\{0\}\)\-P\_\{\\mathrm\{avg\}\}\(z\)\-\\epsilon\_\{P\}\}\{\\epsilon\_\{P\}\}\\right\]\_\{\+\}\\\\ &\+\\left\[\\frac\{\\mathrm\{EOpp\}\(z\)\-d\_\{\\max\}\}\{d\_\{\\max\}\}\\right\]\_\{\+\}\\\\ &\+\\mathbf\{1\}\_\{S\}\\left\[\\frac\{S\(\\rho\)/S\(\\rho\_\{0\}\)\-s\_\{\\max\}\}\{s\_\{\\max\}\}\\right\]\_\{\+\},\\end\{split\}\(14\)where\[t\]\+=max⁡\(0,t\)\[t\]\_\{\+\}=\\max\(0,t\)and𝟏S\\mathbf\{1\}\_\{S\}enables the optional storage bound\. All denominators are positive in the experimental requests\. The diagnostic model carries an infeasible status, and the reported component violations show which requirements remain unsatisfied\. A change to a bound requires a user instruction\.

## VExperiments and Results

### V\-AExperiment Setup

#### V\-A1Dataset and Backbone

The experiments are conducted on Fitzpatrick\-17k\[[33](https://arxiv.org/html/2609.17786#bib.bib18)\], a clinical image dataset for classifying 114 dermatological conditions\. Following previous works, we group Fitzpatrick types 1–3 as Light and types 4–6 as Dark, using skin tone as the sensitive attribute\.

We use a pre\-trained VGG\-11\[[34](https://arxiv.org/html/2609.17786#bib.bib19)\]adapted to112×112112\\times 112inputs and 114 outputs\. The supplied checkpoint is fixed across experiments and is denoted as Vanilla\. Evaluation uses RGB conversion, a shorter\-side resize to 128 pixels, and a center crop, with pixel values scaled to\[0,1\]\[0,1\]\. Fine\-tuning uses random cropping and horizontal flipping\.

#### V\-A2Compression Configurations

The pre\-evaluated candidate set contains5×2×2×2=405\\times 2\\times 2\\times 2=40configurations: five operator sequences, two layer sets, two strengths, andβ∈\{5/9,1\}\\beta\\in\\\{5/9,1\\\}\. Layer set A contains FC\-2, the4096×40964096\\times 4096hidden classifier layer; set B contains both FC\-1 and FC\-2, with FC\-1 having4608×40964608\\times 4096weights in input\-by\-output coordinates\. The output classifier layer remains fixed\. Mild and strong settings use pruning or quantization fractions of 0\.10 and 0\.30, or retained ranks of 1024 and 512 forLL, respectively\. For both ranks, FairLRF sparsifies 20% of factor rows or columns and retains half of their rank coordinates\.

Scoring uses 256 fixed, class\-covering training samples per group and signed diagonal Hessians of the mean cross\-entropy loss\. All configurations receive two fine\-tuning epochs\. ForPPorQQ, the target fraction is reached in two equal increments, each followed by one epoch\. ForLL, both epochs follow factorization and sparsification; each stage of a two\-operator configuration receives one epoch\. This shared fine\-tuning budget provides a common basis for comparing configurations\.

#### V\-A3User Requirements and Search Methods

The preferred trade\-off among accuracy, fairness, and efficiency can be significantly different across various situations\. For simplicity, we consider two straightforward requests fixed before search\. Request U permits a maximumPavgP\_\{\\mathrm\{avg\}\}loss of one percentage point, requires EOpp no worse than Vanilla, and has no storage upper bound\. Request F permits an average precision loss of up to three percentage points, requiresE​O​p​p≤0\.20EOpp\\leq 0\.20, and limits storage to 80% of Vanilla\. Both requests are applied to validation measurements during search\. Feasibility checks use full\-precision measurements\.

To compare with FCA, we adopt random search\[[35](https://arxiv.org/html/2609.17786#bib.bib27)\], a tree\-structured Parzen estimator \(TPE\)\[[36](https://arxiv.org/html/2609.17786#bib.bib20)\], and one\-shot LLM planning as baseline methods, together with Vanilla as their common starting point\. TPE uses a constrained multivariate categorical model with exact acquisition maximization over the unobserved catalogue\. Each search method receives the same three initial observations, one mild configuration from each individual operator family\. These observations count toward the limits of 12 distinct candidates and 120 model minutes per session\. Random search samples without replacement\. One\-shot planning commits to nine further candidates before receiving their outcomes, while FCA adapts subsequent choices and can stop early\. We conduct three runs per search method under each request, yielding 24 search trajectories\.

Both language\-model search methods use GLM\-5\.3\-Flash with the same generation settings and per\-response allowance\. All search methods query the same pre\-evaluated candidate set, with unqueried results kept hidden\. Accumulated model evaluation time includes the queried configuration’s preparation, scoring, fine\-tuning, validation, and I/O\. Language\-model request time is reported separately\. The main comparison combines the completed LLM runs with the corresponding Random/TPE reference trajectories\.

Fig\. 3:Validation outcomes of all 40 compression configurations\. Lower\-right points have higher precision and lower EOpp\. The dashed line is the Vanilla EOpp upper bound used by U\.TABLE I:Within\-family selection under U on validation\. Each compression family contains eight configurations\. Setting gives layer set, strength \(mild or strong\), andβ\\beta\. Selection follows the common rule of minimum storage among feasible observations\. N/A denotes not applicable to Vanilla;\-\-indicates that no feasible configuration is available\.

### V\-BResults of the Compression Operators

Figure[3](https://arxiv.org/html/2609.17786#S5.F3)shows that the operator and its configuration both influence the resulting trade\-off\. Among the 40 configurations, 10 satisfy U: five fromLL, three fromP→QP\\\!\\rightarrow\\\!Q, and two fromL→QL\\\!\\rightarrow\\\!Q\. The individualPPandQQconfigurations fall outside the feasible region under the evaluated fine\-tuning budget\. Feasible configurations therefore span several families and require evaluation at the configuration level\.

Table[I](https://arxiv.org/html/2609.17786#S5.T1)reports the best U\-feasible configuration within each family\. The best candidate in the set isLLon both hidden classifier layers with rank 512 andβ=1\\beta=1\. It hasPavg=0\.5233P\_\{\\mathrm\{avg\}\}=0\.5233,E​O​p​p=0\.2168EOpp=0\.2168, and 73,385,928 inference tensor bytes, corresponding to a 59\.54% storage reduction from Vanilla\. Validation precision increases by 0\.918 percentage points and EOpp decreases by 0\.00829\. These measurements describe the complete configuration after compression and fine\-tuning\.

The composition results illustrate the need for feedback\. Adding quantization to thisLLconfiguration givesPavg=0\.5146P\_\{\\mathrm\{avg\}\}=0\.5146and EOpp=0\.2365=0\.2365at the same tensor storage, placing the resultingL→QL\\\!\\rightarrow\\\!Qcandidate outside U\. Meanwhile,P→QP\\\!\\rightarrow\\\!Qsupplies three feasible configurations even though the evaluated individualPPandQQfamilies supply none\. Thus, both the operator sequence and its settings affect the attainable accuracy–fairness trade\-off\. FCA evaluates the complete configuration before accepting it under the user’s constraints\. The storage ratios ofPP,QQ, andP→QP\\\!\\rightarrow\\\!Qremain 1\.0 because their inference tensors are stored densely in FP32\.

### V\-CComparison of Search Methods

TABLE II:Replay of search methods on the same pre\-evaluated candidate set, with three runs per row\. Storage, precision, and EOpp are coordinate\-wise medians over feasible final selections;vvis the median normalized violation over all runs\. Queries and model minutes are means\. All 24 trajectories in the main comparison complete\.Fig\. 4:Search progress under U and F\. Thin lines show individual runs and thick lines show point\-wise medians; endpoint dots mark stopping times\. The last observation is carried forward for the median after a session stops\. U shows feasible storage and F shows the lowest observed violation\. The horizontal axis reports the accumulated recorded cost of queried model evaluations; language\-model request time is reported separately\.Table[II](https://arxiv.org/html/2609.17786#S5.T2)compares search method outcomes\. Under U, both one\-shot planning and FCA select the best feasible configuration in the candidate set in all three runs, with median storage ratio 0\.4046\. Random and TPE have median storage ratio 0\.5953\. The shared initialization already includes a U\-feasible low\-rank candidate, so this comparison measures improvement in storage among feasible selections\. The model selected by FCA requires 32\.04% less storage than the median storage requirement of models selected by Random or TPE\.

FCA evaluates 7\.33 candidates on average under U, compared with 12 for one\-shot planning, with corresponding accumulated model evaluation times of 13\.66 and 22\.11 minutes\. Its adaptive stopping policy saves evaluations once it has identified a satisfactory trade\-off\. One\-shot planning executes its committed list, so the comparison reflects the complete policies, including their different stopping rules\.

Under F, the lowest EOpp in the entire candidate set is 0\.2168, above the upper bound of 0\.20\. All search methods therefore report an unmet request\. FCA, one\-shot planning, and TPE reach the same median violation of 0\.0839, compared with 0\.1059 for Random\. The 0\.8 storage bound also rules out densePP,QQ, andP→QP\\\!\\rightarrow\\\!Qrepresentations from their known storage requirements\. This case illustrates how structural information guides the search while measured group performance determines the remaining feasibility gap\. Figure[4](https://arxiv.org/html/2609.17786#S5.F4)shows the evolution of the selected storage and violation as observations accumulate\.

TABLE III:Held\-out results for three distinct compression configurations selected before test evaluation\.
### V\-DRepeated Fine\-Tuning and Held\-Out Testing

We evaluate three distinct compression configurations selected from a designated search run and fixed before test evaluation\. Each configuration is compressed and evaluated from the same checkpoint\. The selected configurations areLLwith A/strong/5/95/9, B/mild/1, and B/strong/1\. Matching existing measurements for these configurations are reused\.

Table[III](https://arxiv.org/html/2609.17786#S5.T3)reports held\-out performance\. For B/strong/1, the average precision on the test set is 0\.5357 compared with 0\.5275 for Vanilla, while EOpp is 0\.2562 compared with 0\.2650 for Vanilla\. Meanwhile, it reaches 59\.54% storage reduction\.

### V\-EPlanning Cost

TABLE IV:Language\-model cost summed over three runs per row\. Completion tokens include reported reasoning tokens\. API minutes measure request durations and exclude model execution and the separate online cases\.Table[IV](https://arxiv.org/html/2609.17786#S5.T4)reports the computational cost of planning\. Under U, FCA uses 22 requests and 75,421 completion tokens across the three runs, compared with three requests and 69,653 tokens for one\-shot planning\. The accumulated request durations are 15\.80 and 18\.82 minutes, respectively\. Together with the candidate counts in Table[II](https://arxiv.org/html/2609.17786#S5.T2), these results describe the cost of producing the same final U selection\.

The stricter F request requires more interaction\. FCA uses 35 requests and 154,014 completion tokens, compared with three requests and 67,774 tokens for one\-shot planning, and both reach the same least\-violating model\. The separate request and model\-time measurements characterize the cost of the complete search workflow\.

### V\-FOnline Execution and User Feedback

Two online cases evaluate fresh compression configurations under a six\-candidate limit per case\. In both cases, immediate reselection from existing observations is verified before further model evaluations\.

Under U, the user increases the permitted average precision loss from one to two percentage points\. The final selection isLLwith A/strong/5/95/9, yielding validationPavg=0\.5184P\_\{\\mathrm\{avg\}\}=0\.5184, EOpp=0\.2186=0\.2186, and storage ratio 0\.7225\.

Under F, the user tightens the EOpp upper bound from 0\.20 to 0\.19\. FCA reportsLLwith B/strong/1 as infeasible, with normalized violation 0\.1410\. The report identifies EOpp as the unsatisfied constraint\.

## VIDiscussion

### VI\-AImplications for Compression Selection

The composition results show that an additional operator can change the accuracy–fairness trade\-off of an existing compression configuration\. This makes the complete configuration the appropriate unit of comparison for compression selection\. The U and F results further illustrate how the requested trade\-off determines feasibility: a model accepted under U can remain infeasible under F despite its storage reduction\. FCA makes these differences explicit by associating each recommendation with the measured metrics and the constraints used for selection\.

### VI\-BRepresentation\-Aware Design

The FPGA considerations in Section[I](https://arxiv.org/html/2609.17786#S1)explain why compression should be organized around the resulting representation\. Pruning changes connectivity, quantization constrains arithmetic values, and factorization replaces a layer with smaller matrix operations\. Their suitability depends on the resources and execution patterns supported by the target design\. FCA expresses these distinctions through operator descriptions and compatible operator sequences, and uses model\-level constraints to preserve the user’s accuracy and fairness requirements during selection\. Following the analysis in\[[15](https://arxiv.org/html/2609.17786#bib.bib28)\], such organization makes deployment assumptions explicit and supports extending fairness evaluation to target hardware\.

## VIIConclusion

In this paper, we proposed FairCompressAgent, an agentic framework for selecting fairness\-aware compression configurations under user\-defined requirements\. FCA combines pruning, quantization, and sparse low\-rank factorization through a common operator interface and connects planning to measured model transformations\. Experiments on Fitzpatrick\-17k demonstrate a 59\.54% reduction in inference tensor storage for the model selected on validation under the accuracy\-constrained request, together with explicit reporting of unmet constraints and online adaptation to revised requirements\. The shared evaluation record connects these selections to reproducible model execution and supports continued refinement as the user’s requirements evolve\.

## References

- \[1\]\(2022\)Amazon scraps secret ai recruiting tool that showed bias against women\.InEthics of data and analytics,pp\. 296–299\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[2\]C\. Ferrara, G\. Sellitto, F\. Ferrucci, F\. Palomba, and A\. De Lucia\(2024\)Fairness\-aware machine learning engineering: how far are we?\.Empirical software engineering29\(1\),pp\. 9\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[3\]Z\. Obermeyer, B\. Powers, C\. Vogeli, and S\. Mullainathan\(2019\)Dissecting racial bias in an algorithm used to manage the health of populations\.Science366\(6464\),pp\. 447–453\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[4\]G\. Xu, H\. Fan, and Y\. Shi\(2026\)Group\-conditioned representation modulation for fair skin disease diagnosis\.InMICCAI Student Board 2026 EMERGE Workshop: Empowering Medical Information Computing and Research through Early\-career Guidance and Expertise,Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[5\]G\. Xu, Y\. Duan, Z\. Liu, X\. Li, M\. Jiang, M\. Lemmon, W\. Jin, and Y\. Shi\(2025\)Incorporating rather than eliminating: achieving fairness for skin disease diagnosis through group\-specific experts\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 284–294\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[6\]C\. Chiu, H\. Chung, Y\. Chen, Y\. Shi, and T\. Ho\(2023\)Toward fairness through fair multi\-exit framework for dermatological disease diagnosis\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 97–107\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[7\]S\. Han, H\. Mao, and W\. J\. Dally\(2015\)Deep compression: compressing deep neural networks with pruning, trained quantization and huffman coding\.arXiv preprint arXiv:1510\.00149\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[8\]S\. Han, J\. Pool, J\. Tran, and W\. Dally\(2015\)Learning both weights and connections for efficient neural network\.Advances in neural information processing systems28\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[9\]J\. Luo, J\. Wu, and W\. Lin\(2017\)Thinet: A filter level pruning method for deep neural network compression\.InProceedings of the IEEE international conference on computer vision,pp\. 5058–5066\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[10\]B\. Jacob, S\. Kligys, B\. Chen, M\. Zhu, M\. Tang, A\. Howard, H\. Adam, and D\. Kalenichenko\(2018\)Quantization and training of neural networks for efficient integer\-arithmetic\-only inference\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\. 2704–2713\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[11\]S\. Zhou, Y\. Wu, Z\. Ni, X\. Zhou, H\. Wen, and Y\. Zou\(2016\)Dorefa\-net: training low bitwidth convolutional neural networks with low bitwidth gradients\.arXiv preprint arXiv:1606\.06160\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[12\]S\. Hooker, N\. Moorosi, G\. Clark, S\. Bengio, and E\. Denton\(2020\)Characterising bias in compressed models\.arXiv preprint arXiv:2010\.03058\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p1.1)\.
- \[13\]C\. Zhang, P\. Li, G\. Sun, Y\. Guan, B\. Xiao, and J\. Cong\(2015\)Optimizing fpga\-based accelerator design for deep convolutional neural networks\.InProceedings of the 2015 ACM/SIGDA international symposium on field\-programmable gate arrays,pp\. 161–170\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p2.1)\.
- \[14\]Y\. Umuroglu, N\. J\. Fraser, G\. Gambardella, M\. Blott, P\. Leong, M\. Jahre, and K\. Vissers\(2017\)Finn: a framework for fast, scalable binarized neural network inference\.InProceedings of the 2017 ACM/SIGDA international symposium on field\-programmable gate arrays,pp\. 65–74\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p2.1)\.
- \[15\]Y\. Guo, Z\. Yan, X\. Yu, Q\. Kong, J\. Xie, K\. Luo, D\. Zeng, Y\. Wu, Z\. Jia, and Y\. Shi\(2024\)Hardware design and the fairness of a neural network\.Nature Electronics7\(8\),pp\. 714–723\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p2.1),[§VI\-B](https://arxiv.org/html/2609.17786#S6.SS2.p1.1)\.
- \[16\]S\. S\. Mugdho, Y\. Guo, E\. G\. Rogers, W\. Zhao, Y\. Shi, and C\. Wang\(2025\)FairXbar: improving the fairness of deep neural networks with non\-ideal in\-memory computing hardware\.In2025 Design, Automation & Test in Europe Conference \(DATE\),pp\. 1–7\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1)\.
- \[17\]Y\. Sheng, J\. Yang, J\. Li, J\. Alaina, X\. Xu, Y\. Shi, J\. Hu, W\. Jiang, and L\. Yang\(2024\)Data\-algorithm\-architecture co\-optimization for fair neural networks on skin lesion dataset\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 153–163\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1)\.
- \[18\]R\. Qin, Y\. Hu, Z\. Yan, J\. Xiong, A\. Abbasi, and Y\. Shi\(2024\)Fl\-nas: towards fairness of nas for resource constrained devices via large language models\.In2024 29th Asia and South Pacific Design Automation Conference \(ASP\-DAC\),pp\. 429–434\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1)\.
- \[19\]Y\. Sheng, J\. Yang, Y\. Wu, K\. Mao, Y\. Shi, J\. Hu, W\. Jiang, and L\. Yang\(2022\)The larger the fairer? small neural networks can achieve fairness for edge devices\.InProceedings of the 59th ACM/IEEE Design Automation Conference,pp\. 163–168\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1)\.
- \[20\]Y\. Wu, D\. Zeng, X\. Xu, Y\. Shi, and J\. Hu\(2022\)Fairprune: achieving fairness through pruning for dermatological disease diagnosis\.InMedical Image Computing and Computer Assisted Intervention–MICCAI 2022: 25th International Conference, Singapore, September 18–22, 2022, Proceedings, Part I,pp\. 743–753\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1),[§III\-A](https://arxiv.org/html/2609.17786#S3.SS1.p1.1),[§III\-B](https://arxiv.org/html/2609.17786#S3.SS2.p1.1)\.
- \[21\]Y\. Guo, Z\. Jia, J\. Hu, and Y\. Shi\(2024\)FairQuantize: achieving fairness through weight quantization for dermatological disease diagnosis\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 329–338\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1),[§III\-A](https://arxiv.org/html/2609.17786#S3.SS1.p1.1),[§III\-B](https://arxiv.org/html/2609.17786#S3.SS2.p2.1)\.
- \[22\]S\. Swaminathan, D\. Garg, R\. Kannan, and F\. Andres\(2020\)Sparse low rank factorization for deep neural network compression\.Neurocomputing398,pp\. 185–196\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1),[§III\-C](https://arxiv.org/html/2609.17786#S3.SS3.p1.2)\.
- \[23\]Y\. Guo, J\. Xia, and Y\. Shi\(2025\)FairLRF: achieving fairness through sparse low rank factorization\.arXiv preprint arXiv:2511\.16549\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1),[§III\-A](https://arxiv.org/html/2609.17786#S3.SS1.p1.1),[§III\-C](https://arxiv.org/html/2609.17786#S3.SS3.p1.2)\.
- \[24\]Q\. Kong, C\. Chiu, D\. Zeng, Y\. Chen, T\. Ho, J\. Hu, and Y\. Shi\(2024\)Achieving fairness through channel pruning for dermatological disease diagnosis\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 24–34\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p3.1)\.
- \[25\]T\. Wang, K\. Wang, H\. Cai, J\. Lin, Z\. Liu, H\. Wang, Y\. Lin, and S\. Han\(2020\)Apq: joint search for network architecture, pruning and quantization policy\.In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 2075–2084\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p4.1)\.
- \[26\]T\. Liu, N\. Astorga, N\. Seedat, and M\. van der Schaar\(2024\)Large language models to enhance bayesian optimization\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 31252–31284\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p4.1)\.
- \[27\]Y\. Dai, L\. Zhang, F\. Luo, M\. Chowdhury, and Y\. Wu\(2025\)FairAgent: democratizing fairness\-aware machine learning with llm\-powered agents\.In2025 IEEE International Conference on Data Mining Workshops \(ICDMW\),pp\. 2597–2601\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p4.1)\.
- \[28\]S\. Jafari, A\. Sarkar, M\. Bilwal, and A\. Jannesari\(2025\)ProfilingAgent: profiling\-guided agentic reasoning for adaptive model optimization\.arXiv preprint arXiv:2509\.05584\.Cited by:[§I](https://arxiv.org/html/2609.17786#S1.p4.1)\.
- \[29\]M\. Hardt, E\. Price, and N\. Srebro\(2016\)Equality of opportunity in supervised learning\.Advances in neural information processing systems29\.Cited by:[§II\-B](https://arxiv.org/html/2609.17786#S2.SS2.p1.1)\.
- \[30\]Y\. LeCun, J\. Denker, and S\. Solla\(1989\)Optimal brain damage\.Advances in neural information processing systems2\.Cited by:[§III\-A](https://arxiv.org/html/2609.17786#S3.SS1.p1.1)\.
- \[31\]A\. Zhou, A\. Yao, Y\. Guo, L\. Xu, and Y\. Chen\(2017\)Incremental network quantization: towards lossless cnns with low\-precision weights\.arXiv preprint arXiv:1702\.03044\.Cited by:[§III\-B](https://arxiv.org/html/2609.17786#S3.SS2.p2.2)\.
- \[32\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2022\)React: synergizing reasoning and acting in language models\.arXiv preprint arXiv:2210\.03629\.Cited by:[§IV\-A](https://arxiv.org/html/2609.17786#S4.SS1.p1.1)\.
- \[33\]M\. Groh, C\. Harris, L\. Soenksen, F\. Lau, R\. Han, A\. Kim, A\. Koochek, and O\. Badri\(2021\)Evaluating deep neural networks trained on clinical images in dermatology with the fitzpatrick 17k dataset\.In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops \(CVPRW\),pp\. 1820–1828\.Cited by:[§V\-A1](https://arxiv.org/html/2609.17786#S5.SS1.SSS1.p1.1)\.
- \[34\]K\. Simonyan and A\. Zisserman\(2014\)Very deep convolutional networks for large\-scale image recognition\.arXiv preprint arXiv:1409\.1556\.Cited by:[§V\-A1](https://arxiv.org/html/2609.17786#S5.SS1.SSS1.p2.1)\.
- \[35\]J\. Bergstra and Y\. Bengio\(2012\)Random search for hyper\-parameter optimization\.\.Journal of machine learning research13\(2\)\.Cited by:[§V\-A3](https://arxiv.org/html/2609.17786#S5.SS1.SSS3.p2.1)\.
- \[36\]J\. Bergstra, R\. Bardenet, Y\. Bengio, and B\. Kégl\(2011\)Algorithms for hyper\-parameter optimization\.Advances in neural information processing systems24\.Cited by:[§V\-A3](https://arxiv.org/html/2609.17786#S5.SS1.SSS3.p2.1)\.