#!/usr/bin/env python3 """Draw corpus statistics and within-paper comparisons; never overwrite files. Requires matplotlib and numpy. All plotted observations are listed in plot_data.json. """ from pathlib import Path import json import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt from matplotlib.colors import ListedColormap from matplotlib.patches import Patch import numpy as np HERE=Path(__file__).resolve().parent FIGURES=HERE/'figures' INK='#193047'; MUTED='#627486'; TEAL='#187f79'; BLUE='#6685bd' ORANGE='#c47b36'; GREY='#aab4bf'; PALE='#f0f3f5'; PAPER='#ffffff' plt.rcParams.update({ 'font.family':'DejaVu Sans','font.size':10,'text.color':INK, 'axes.labelcolor':INK,'xtick.color':MUTED,'ytick.color':MUTED, 'axes.titleweight':'bold','axes.titlesize':12,'axes.labelsize':10, 'axes.edgecolor':'#cbd2d9','axes.spines.top':False,'axes.spines.right':False, 'figure.facecolor':PAPER,'axes.facecolor':PAPER,'savefig.facecolor':PAPER, 'grid.color':'#e7ebee','grid.linewidth':0.7,'axes.axisbelow':True, 'svg.fonttype':'none','savefig.dpi':185, }) TRIADIC=[ # name, state MB, parameters M, recall %, PG19 16K-64K PPL, group, PDF page ('GDN base',6.3,380.9,26.2,14.15,'triadic',6), ('Larger heads (2x)',12.6,380.7,26.6,14.17,'alternative',6), ('Grouped values (2x)',12.6,378.2,27.1,14.12,'alternative',6), ('Wider values (2x)',12.6,377.9,27.4,14.14,'alternative',6), ('More heads (2x)',12.6,381.6,27.5,14.28,'alternative',6), ('Triadic E=2',12.6,381.9,28.4,13.95,'triadic',6), ('Larger heads (4x)',25.2,380.6,28.5,14.24,'alternative',6), ('Grouped values (4x)',25.2,382.4,27.3,14.36,'alternative',6), ('Wider values (4x)',25.2,381.3,27.5,14.44,'alternative',6), ('Triadic E=4',25.2,383.0,31.1,13.79,'triadic',6), ('Triadic E=8, scratch',50.3,385.4,33.1,13.74,'triadic',7), ('Triadic E=8, upcycled',50.3,385.4,29.7,13.86,'upcycled',7), ] LATTICE=[[16,15.52],[32,15.26],[64,15.02],[96,14.89],[128,14.84],[160,14.76],[192,14.71]] def save(fig,name): for suffix in ('.png','.svg'): p=FIGURES/(name+suffix) if p.exists(): raise FileExistsError(f'Refusing to overwrite {p}') for suffix in ('.png','.svg'): fig.savefig(FIGURES/(name+suffix),bbox_inches='tight',pad_inches=0.22) plt.close(fig) def landscape(corpus): fig=plt.figure(figsize=(13.2,7.5)) gs=fig.add_gridspec(1,2,width_ratios=[2.3,1],wspace=.19) ax=fig.add_subplot(gs[0,0]); ax2=fig.add_subplot(gs[0,1]) lookup={p['name']:p for p in corpus['papers']} labels=[]; direct=[];nearby=[] for group in corpus['groups']: labels.append(group['en']) direct.append(sum(lookup[n]['tier']=='A' for n in group['members'])) nearby.append(sum(lookup[n]['tier']=='B' for n in group['members'])) y=np.arange(len(labels)) ax.barh(y,direct,color=TEAL,height=.64,label='A: directly related (45)') ax.barh(y,nearby,left=direct,color=BLUE,height=.64,label='B: adjacent mechanisms (23)') for yy,a,b in zip(y,direct,nearby): ax.text(a+b+.22,yy,str(a+b),va='center',fontweight='bold',fontsize=11) ax.set_yticks(y,labels);ax.invert_yaxis();ax.set_xlim(0,13) ax.set_xticks(range(0,14,2));ax.set_xlabel('Unique papers in this library') ax.grid(axis='x');ax.tick_params(axis='y',length=0) ax.set_title('68 related papers across nine primary categories',loc='left',pad=18) ax.legend(loc='lower left',bbox_to_anchor=(0,-.24),ncol=1,frameon=False,fontsize=10) ax2.pie([45,23,67],colors=[TEAL,BLUE,'#dfe4e9'],startangle=90,counterclock=False, wedgeprops={'width':.25,'edgecolor':'white','linewidth':3}) ax2.text(0,.08,'135',ha='center',fontsize=34,fontweight='bold') ax2.text(0,-.17,'unique papers',ha='center',color=MUTED,fontsize=11) ax2.set_title('Local library snapshot',pad=15) ax2.text(0,-1.35,'45 direct + 23 adjacent\n67 outside the mechanism count', ha='center',va='top',fontsize=11,linespacing=1.7) fig.suptitle('Linear attention: a dense landscape of memory designs',x=.02,y=.99, ha='left',fontsize=17,fontweight='bold') fig.text(.02,.005,'2026-10-20 | Local corpus, not a global publication census. ' 'Primary categories are exclusive; paper versions and note revisions are deduplicated.',fontsize=9,color=MUTED) fig.subplots_adjust(left=.265,right=.98,top=.87,bottom=.2) save(fig,'01-corpus-landscape') def mechanism_map(data): rows=data['rows']; cols=data['columns']; matrix=np.zeros((len(rows),len(cols))) for i,row in enumerate(rows): for j in row['features']: matrix[i,j]=1 if row['tier']=='A' else 2 fig,ax=plt.subplots(figsize=(12.2,13.8)) ax.imshow(matrix,cmap=ListedColormap([PALE,TEAL,BLUE]),vmin=0,vmax=2, aspect='auto',interpolation='nearest') xlabels=['Content\ncapacity +','Auxiliary\nstate +','Routing /\nslots','Write /\nerase', 'Time /\nphase','Readout','Training /\ninput','Storage /\nI/O','Explicit\nKV'] ax.set_xticks(np.arange(len(cols)),xlabels,fontsize=10) ax.xaxis.tick_top();ax.tick_params(axis='x',pad=10,length=0) ax.set_yticks(np.arange(len(rows)),[r['name'] for r in rows],fontsize=10) ax.tick_params(axis='y',pad=10,length=0) ax.set_xticks(np.arange(-.5,len(cols),1),minor=True) ax.set_yticks(np.arange(-.5,len(rows),1),minor=True) ax.grid(which='minor',color='white',linewidth=3) ax.tick_params(which='minor',bottom=False,left=False) for spine in ax.spines.values():spine.set_visible(False) for i in range(len(rows)): for j in range(len(cols)): if matrix[i,j]:ax.text(j,i,'•',ha='center',va='center',fontsize=12,color='white') fig.suptitle('Many papers act on neighboring parts of the same memory lifecycle', x=.02,y=.995,ha='left',fontsize=16,fontweight='bold') fig.text(.02,.961,'34 representative works | Qualitative intervention map; ' 'colored cells are not performance scores.',fontsize=11,color=MUTED) fig.legend(handles=[Patch(facecolor=TEAL,label='A: direct'),Patch(facecolor=BLUE,label='B: adjacent'), Patch(facecolor=PALE,label='Not the highlighted intervention')], loc='lower left',bbox_to_anchor=(.17,.035),ncol=3,frameon=False,fontsize=10) fig.text(.02,.018,'Extra content capacity and auxiliary memory are separate. ' 'Flags describe the design relative to its base path; they do not assert equal-resource gains.',fontsize=9,color=MUTED) fig.subplots_adjust(left=.21,right=.99,top=.87,bottom=.10) save(fig,'02-mechanism-overlap') def triadic(): fig,axes=plt.subplots(1,2,figsize=(12.6,5.9)) selected=[r for r in TRIADIC if r[5]=='triadic'];alts=[r for r in TRIADIC if r[5]=='alternative'] up=[r for r in TRIADIC if r[5]=='upcycled'][0] for ax,idx,title,ylabel,ref in zip(axes,[3,4],['Recall improves, with a remaining gap','Long-context prediction improves'], ['Six-task recall average (%)','PG19 16K–64K perplexity'],[41.6,13.92]): ax.scatter([r[1] for r in alts],[r[idx] for r in alts],s=49,color=GREY, edgecolor='white',linewidth=.6,zorder=3,label='Other enlargements (Table 1)') ax.plot([r[1] for r in selected],[r[idx] for r in selected],'-o',color=TEAL, linewidth=2.6,markersize=7,label='Triadic, trained from scratch',zorder=4) ax.scatter(up[1],up[idx],marker='D',s=78,color=ORANGE,edgecolor='white', zorder=5,label='E=8, upcycled') ax.axhline(ref,color=BLUE,linestyle=(0,(4,3)),linewidth=1.5, label='Transformer (growing KV; reference)') for r in selected: label=f'{r[idx]:.1f}' if idx==3 else f'{r[idx]:.2f}' offset=(0,11) if idx==3 else (0,-18) ax.annotate(label,(r[1],r[idx]),xytext=offset,textcoords='offset points', ha='center',color=TEAL,fontsize=10,fontweight='bold') ax.annotate(f'{up[idx]:.1f}' if idx==3 else f'{up[idx]:.2f}',(up[1],up[idx]), xytext=(-7,12) if idx==3 else (-8,-6),textcoords='offset points', ha='right',color=ORANGE,fontsize=10) ax.set_xscale('log',base=2);ax.set_xticks([6.3,12.6,25.2,50.3],['6.3','12.6','25.2','50.3']) ax.set_xlim(5.2,61);ax.set_xlabel('Content state (MB; 2 bytes / entry)') ax.set_ylabel(ylabel);ax.set_title(title,loc='left',pad=15);ax.grid(alpha=.8) axes[0].set_ylim(23,44);axes[0].set_yticks([25,30,35,40]) axes[1].set_ylim(13.57,14.57) handles,labels=axes[0].get_legend_handles_labels() fig.legend(handles,labels,loc='lower center',bbox_to_anchor=(.5,.065),ncol=2,frameon=False,fontsize=9.5) fig.suptitle('Capacity is useful; its organization and training also matter', x=.02,y=.99,ha='left',fontsize=17,fontweight='bold') fig.text(.02,.018,'Redrawn from Triadic v1, Tables 1–2 (PDF pp.6–7). ' '400M-class GDN; means of 3 seeds. Parameters vary slightly; no invented error bars.',fontsize=9,color=MUTED) fig.subplots_adjust(left=.07,right=.98,wspace=.26,top=.81,bottom=.29) save(fig,'03-triadic-capacity') def utilization(): fig,axes=plt.subplots(1,2,figsize=(12.6,5.9),gridspec_kw={'wspace':.29}) ax=axes[0];m,ppl=np.array(LATTICE).T ax.plot(m,ppl,'-o',color=TEAL,linewidth=2.6,markersize=6,label='Lattice') ax.scatter([64],[15.60],marker='D',s=80,color=ORANGE,label='TTT, 64 slots',zorder=5) ax.scatter([64],[16.35],marker='s',s=60,color=BLUE,label='DeltaNet, 64 slots',zorder=5) ax.annotate('16 slots: 15.52',(16,15.52),xytext=(9,14),textcoords='offset points',color=TEAL,fontweight='bold') ax.annotate('15.60',(64,15.60),xytext=(8,7),textcoords='offset points',color=ORANGE) ax.set_xscale('log',base=2);ax.set_xticks([16,32,64,128,192],['16','32','64','128','192']) ax.set_xlim(13,230);ax.set_ylim(14.45,16.65) ax.set_ylabel('Validation perplexity (lower is better)');ax.set_xlabel('Compression memory slots m') ax.set_title('Lattice: memory efficiency within one paper',loc='left',pad=14) ax.grid(alpha=.8);ax.legend(frameon=False,loc='upper right',fontsize=9) ax.text(0,-.23,'110M-class models, FineWeb-Edu\nTable 3 (p.10), Figure 6 (p.23)', transform=ax.transAxes,fontsize=9,color=MUTED,linespacing=1.6) ax=axes[1] bars=ax.bar([0,1],[8.8,97.4],width=.55,color=[GREY,TEAL]) for bar,value in zip(bars,[8.8,97.4]): ax.text(bar.get_x()+bar.get_width()/2,value+2,f'{value:.1f}%',ha='center', va='bottom',fontweight='bold',fontsize=13) ax.set_xticks([0,1],['Window 512\nGlobal 0','Window 256\nGlobal 256 (SRE)']) ax.set_ylim(0,111);ax.set_yticks([0,25,50,75,100]);ax.grid(axis='y') ax.set_ylabel('4K S-NIAH-1 accuracy (%)');ax.set_title('LoLA: same nominal explicit-cache budget',loc='left',pad=14) ax.text(.04,.77,'Same weights\n512 nominal KV slots',transform=ax.transAxes,ha='left', color=MUTED,fontsize=11,linespacing=1.6) ax.text(0,-.23,'LoLCATs-8B+ vs LoLA-8B\nLoLA v3, Table 1 (PDF pp.7–8)', transform=ax.transAxes,fontsize=9,color=MUTED,linespacing=1.6) fig.suptitle('Utilization requires a resource-aware comparison',x=.02,y=.99, ha='left',fontsize=17,fontweight='bold') fig.text(.02,.018,'Each panel is a separate within-paper comparison. ' 'The two panels do not share a benchmark or establish a cross-paper ranking.',fontsize=9,color=MUTED) fig.subplots_adjust(left=.07,right=.98,top=.81,bottom=.27) save(fig,'04-utilization-evidence') def main(): corpus=json.loads((HERE/'corpus_inventory.json').read_text()) mechanism=json.loads((HERE/'mechanism_map.json').read_text()) plot_data={ 'scope':'Measured results are compared only within the named paper and protocol.', 'triadic':{ 'source_pdf':'paper-readings/Triadic-2609.36529/v1/source/paper.pdf', 'source_tex':['tables/main_400m_combined.tex','tables/upcycling.tex'], 'source_tex_root':'paper-readings/Triadic-2609.36529/v1/source/tex/files', 'columns':['model','state_mb_2bytes_per_entry','parameters_million','recall_percent','pg19_16k_64k_ppl','group','pdf_page'], 'rows':TRIADIC,'transformer_reference':{'recall_percent':41.6,'pg19_16k_64k_ppl':13.92,'pdf_page':6}, 'protocol':'400M-class GDN; three-seed mean; 20B-token pretraining + long-context extension as in paper; E=8 upcycled is a separate training condition.', }, 'lattice':{ 'source_pdf':'paper-readings/Lattice-2504.05646/v2/source/paper.pdf', 'source_tex':'paper-readings/Lattice-2504.05646/v2/source/tex/files/FigureTable/ablation_m.tex', 'source_location':'Table 3 PDF p.10; Figure 6 PDF p.23', 'columns':['memory_slots','validation_ppl'],'lattice_rows':LATTICE, 'baselines':{'TTT':[64,15.60],'DeltaNet':[64,16.35]}, 'protocol':'110M-class models on FineWeb-Edu, configurations as reported; memory slot count is not full-model bytes.', }, 'lola':{ 'source_pdf':'paper-readings/LoLA-2505.23666/v3/source/paper.pdf', 'source_location':'Table 1 PDF pp.7–8', 'protocol':'Same model weights; 4K S-NIAH-1; 512 nominal explicit KV slots; not an assertion of identical peak VRAM.', 'rows':[{'model':'LoLCATs-8B+','window':512,'global':0,'accuracy_percent':8.8}, {'model':'LoLA-8B','window':256,'global':256,'accuracy_percent':97.4}], }, 'corpus':{'input':'corpus_inventory.json','total':135,'direct':45,'adjacent':23,'outside_count':67}, 'mechanism_map':{'input':'mechanism_map.json','representative_rows':len(mechanism['rows']), 'meaning':'Qualitative intervention classification by this report, not empirical scores.'}, } with (HERE/'plot_data.json').open('x',encoding='utf-8') as f: json.dump(plot_data,f,ensure_ascii=False,indent=2);f.write('\n') landscape(corpus);mechanism_map(mechanism);triadic();utilization() print('Created 4 PNG figures, 4 SVG figures, and plot_data.json.') if __name__=='__main__':main()