Structuring Agentic Research with Evidence Graphs

# Illustrative sketch (not from the paper)
import torch
from collections import defaultdict

def decompose_and_fetch(q):
    subqs=["sub1","sub2"]; seeds=[]
    for _ in subqs: seeds+=["paperA","paperB"]
    return seeds

def citation_neighbors(p): return ["c1","c2"]

def entailment(c,a): return torch.rand(1).item()

def build_graph(seeds):
    g=defaultdict(dict)
    for s in seeds:
        for n in citation_neighbors(s):
            sc=entailment("claim","abs")
            if sc>0.5: g[s][n]=sc
    return g

def random_walk(g,steps=10,alpha=0.85):
    scores={n:0 for n in g}; cur=next(iter(g))
    for _ in range(steps):
        scores[cur]+=1
        if torch.rand(1).item()<alpha and g[cur]: cur=next(iter(g[cur]))
        else: cur=next(iter(g))
    return sorted(scores,key=scores.get,reverse=True)[:5]

def crase(q):
    return random_walk(build_graph(decompose_and_fetch(q)))

print(crase("example query"))

// Illustrative sketch (not from the paper)
const torch=require('torch'); // placeholder
function decomposeAndFetch(q){
  const subqs=["sub1","sub2"], seeds=[];
  subqs.forEach(()=>seeds.push("paperA","paperB"));
  return seeds;
}
function citationNeighbors(p){return["c1","c2"];}
function entailment(c,a){return Math.random();}
function buildGraph(seeds){
  const g={};
  seeds.forEach(s=>citationNeighbors(s).forEach(n=>{const sc=entailment("claim","abs");if(sc>0.5){if(!g[s])g[s]={};g[s][n]=sc;}}));
  return g;
}
function randomWalk(g,steps=10,alpha=0.85){
  const scores={};Object.keys(g).forEach(n=>scores[n]=0);
  let cur=Object.keys(g)[0];
  for(let i=0;i<steps;i++){
    scores[cur]++;
    if(Math.random()<alpha && Object.keys(g[cur]).length){cur=Object.keys(g[cur])[0];}else{cur=Object.keys(g)[0];}
  }
  return Object.entries(scores).sort((a,b)=>b[1]-a[1]).slice(0,5).map(e=>e[0]);
}
function crase(q){return randomWalk(buildGraph(decomposeAndFetch(q)));}
console.log(crase("example query"));