Phase-Locked Dynamic Sampling and Frequency-Domain Adaptation: Attractor-Guided Inference and Spectral Sparsity in LLMs

{"Large":[0],"Language":[1],"Models":[2],"(LLMs)":[3],"frequently":[4],"suffer":[5],"from":[6],"reasoning":[7],"degradation":[8],"during":[9,16],"complex":[10],"autoregressive":[11],"generation":[12],"and":[13,30,54,139,145],"parameter":[14],"redundancy":[15],"fine-tuning.":[17],"In":[18],"this":[19],"paper,":[20],"we":[21],"propose":[22],"two":[23],"synergistic":[24],"systems:":[25],"Phase-Locked":[26],"Dynamic":[27,31,88],"Sampling":[28],"(PLDS)":[29],"F-LoRA":[32,89],"(Fourier-LoRA).":[33],"PLDS":[34,76],"acts":[35],"as":[36,130],"an":[37,119,131],"inference":[38],"controller":[39],"that":[40,75],"models":[41],"internal":[42],"hidden":[43],"state":[44],"activation":[45],"trajectories":[46],"via":[47,124],"a":[48,78,100,109],"non-linear":[49],"chaotic":[50],"attractor":[51],"(Rossler":[52],"system)":[53],"dynamically":[55],"modulates":[56],"the":[57,63,83,94],"sampling":[58],"temperature":[59],"T(t)":[60],"based":[61],"on":[62,82],"derivative":[64],"of":[65,115],"phase":[66],"divergence":[67],"velocity":[68],"St":[69],"=":[70],"|d(delta_phi)/dt|.":[71],"Empirical":[72],"evaluations":[73],"show":[74],"achieves":[77],"+2.7%":[79],"Pass@1":[80],"improvement":[81],"GSM8K":[84],"benchmark":[85,146],"(reaching":[86],"79.1%).":[87],"provides":[90],"parameter-efficient":[91],"fine-tuning":[92],"in":[93],"frequency":[95],"domain":[96],"(2D-DCT),":[97],"guided":[98],"by":[99],"Gumbel-Sigmoid":[101],"Straight-Through":[102],"Estimator":[103],"(STE)":[104],"gating":[105],"mechanism.":[106],"We":[107],"reveal":[108],"critical":[110],"spectral":[111,134],"phenomenon:":[112],"pruning":[113],"75-85%":[114],"high-frequency":[116],"parameters":[117],"(with":[118],"optimal":[120],"threshold":[121],"at":[122,151],"80%)":[123],"explicit":[125],"sparsity":[126],"regularization":[127],"(lambda_sparsity)":[128],"functions":[129],"effective":[132],"low-pass":[133],"filter,":[135],"suppressing":[136],"structural":[137],"noise":[138],"overfitting.":[140],"\\"The":[141],"official":[142],"PyTorch":[143],"implementation":[144],"code":[147],"are":[148],"publicly":[149],"available":[150],"https://github.com/yana-shlyakhova/plds-f-lora-llm":[152],"\\"":[153]}