adu2021 REDSearcher: Scalable Framework for Long-Horizon Search Agents
Problem Context
Language models for information-seeking (web search, retrieval) face extreme sparsity: high-quality search trajectories are rare, multi-step reasoning is expensive, and real-world API costs prohibit extensive exploration. REDSearcher addresses this by (1) generating diverse, challenging synthetic tasks, (2) training fundamental skills in controlled environments, (3) scaling to real interaction only after skills solidify.
Core Concept
REDSearcher operates in three phases: (1) dual-constrained task synthesis creating complex search problems, (2) atomic skill learning (grounding, planning) in simulation, (3) hierarchical trajectory training with environment interaction. This staged approach reduces expensive real-world costs while maintaining quality.
Architecture Overview
- Dual constraints: Control task complexity (graph treewidth) and information dispersion
- Tool-augmented synthesis: Convert facts into tool-resolvable constraints (API calls)
- Atomic skills: Intent grounding, hierarchical planning, fact composition
- Simulation environment: Local search environment with millions of documents
- Staged training: Skills → hierarchical planning → environment interaction
- Long-horizon support: 30+ step trajectories with explicit reasoning
Implementation
Step 1: Dual-constrained task synthesis
import networkx as nx
from typing import Dict, List, Tuple, Set
import random
class DualConstraintedTaskSynthesis:
"""Generate search tasks with controlled complexity."""
def __init__(
self,
document_corpus: List[str],
entity_graph: nx.DiGraph,
max_treewidth: int = 5
):
self.corpus = document_corpus
self.entity_graph = entity_graph
self.max_treewidth = max_treewidth
def compute_treewidth(self, subgraph: nx.DiGraph) -> int:
"""Estimate graph treewidth (NP-hard; use heuristic)."""
# Simplified: use degree-based approximation
if len(subgraph.nodes) == 0:
return 0
degrees = [subgraph.degree(n) for n in subgraph.nodes]
return max(degrees) if degrees else 0
def synthesize_query_with_complexity(
self,
target_treewidth: int = 3,
information_dispersion: float = 0.7
) -> Dict:
"""
Generate query with target topological complexity.
Args:
target_treewidth: Desired graph complexity (higher = harder routing)
information_dispersion: Fraction of facts in different documents (0-1)
Returns:
task: {question, gold_facts, required_searches, optimal_path}
"""
# Sample entities to connect
num_entities = min(target_treewidth + 2, len(self.entity_graph.nodes))
selected_entities = random.sample(list(self.entity_graph.nodes), num_entities)
# Create subgraph and verify treewidth
subgraph = self.entity_graph.subgraph(selected_entities)
actual_treewidth = self.compute_treewidth(subgraph)
if actual_treewidth > self.max_treewidth:
# Simplify by removing high-degree nodes
high_degree_nodes = [n for n in subgraph.nodes if subgraph.degree(n) > 3]
selected_entities = [e for e in selected_entities if e not in high_degree_nodes]
# Generate gold facts from entity connections
gold_facts = []
for source, target in subgraph.edges:
relation = self.entity_graph.edges[source, target].get('relation', 'connects_to')
gold_facts.append(f"{source} {relation} {target}")
# Distribute facts across documents (information dispersion)
facts_per_doc = max(1, len(gold_facts) // max(1, int(len(self.corpus) * information_dispersion)))
document_assignments = {}
for doc_idx in range(len(self.corpus)):
assigned_facts = gold_facts[
doc_idx * facts_per_doc:(doc_idx + 1) * facts_per_doc
]
if assigned_facts:
document_assignments[doc_idx] = assigned_facts
# Generate question requiring all facts
question = self._generate_question(selected_entities, gold_facts)
return {
'question': question,
'gold_facts': gold_facts,
'document_assignments': document_assignments,
'required_entities': selected_entities,
'treewidth': actual_treewidth,
'dispersion': information_dispersion
}
def _generate_question(self, entities: List[str], facts: List[str]) -> str:
"""Generate natural language question from entities and facts."""
if len(entities) >= 2:
return f"What is the connection between {entities[0]} and {entities[-1]}? " \
f"List all intermediate steps."
return f"Describe the relationships for: {', '.join(entities)}"
Step 2: Tool-augmented learning
class ToolAugmentedConstraintRepresentation:
"""Convert facts into tool-resolvable constraints."""
def __init__(self, available_tools: Dict[str, callable]):
"""
Args:
available_tools: Dict of tool_name -> tool_function
e.g., {search_web, get_entity_facts, route_query}
"""
self.tools = available_tools
def convert_fact_to_tool_constraint(
self,
fact: str
) -> Dict:
"""
Convert fact into tool call constraint.
Example:
Fact: "Alice works at Acme Corp"
Constraint: {tool: "get_entity_facts", entity: "Alice", expected: "Acme Corp"}
"""
# Parse fact structure (simplified; would use NER in practice)
parts = fact.split()
if "connects" in fact.lower() or "relationship" in fact.lower():
# Entity-relation-entity triple
entity1, entity2 = parts[0], parts[-1]
constraint = {
'type': 'entity_relationship',
'entity_1': entity1,
'entity_2': entity2,
'required_tool': 'get_entity_relationship',
'expected_relation': fact
}
else:
# Property fact
entity = parts[0]
property_val = ' '.join(parts[1:])
constraint = {
'type': 'entity_property',
'entity': entity,
'property': property_val,
'required_tool': 'get_entity_facts',
'expected_value': property_val
}
return constraint
def create_tool_sequence(self, task: Dict) -> List[Dict]:
"""
Create sequence of tool calls needed to resolve task.
Returns:
tool_sequence: List of {tool_name, arguments, expected_output}
"""
tool_sequence = []
for fact in task['gold_facts']:
constraint = self.convert_fact_to_tool_constraint(fact)
tool_call = {
'tool': constraint['required_tool'],
'arguments': {
k: v for k, v in constraint.items()
if k not in ['type', 'required_tool', 'expected_relation', 'expected_value']
},
'expected_output': constraint.get('expected_relation', constraint.get('expected_value'))
}
tool_sequence.append(tool_call)
return tool_sequence
Step 3: Atomic skill learning in simulation
class AtomicSkillTrainer:
"""Train fundamental search skills in simulation."""
def __init__(
self,
model,
optimizer,
simulated_environment: dict
):
self.model = model
self.optimizer = optimizer
self.env = simulated_environment
def train_intent_grounding(
self,
tasks: List[Dict],
num_steps: int = 1000
) -> float:
"""
Train skill: ground natural language intent into API calls.
Task: Given question, predict correct tool and arguments.
"""
total_loss = 0.0
for step in range(num_steps):
task = random.choice(tasks)
question = task['question']
# Ground intent: predict first tool call
tool_sequence = task['tool_sequence']
if not tool_sequence:
continue
first_tool = tool_sequence[0]
# Generate grounding
grounding_prompt = f"Question: {question}\n\nPredicted next search:"
predicted_grounding, log_probs = self.model.generate_with_logprobs(
grounding_prompt, max_tokens=100
)
# Reward: does predicted grounding match first tool?
is_correct = self._match_grounding(
predicted_grounding,
first_tool['tool']
)
reward = 1.0 if is_correct else 0.0
# Policy gradient update
loss = -log_probs.mean() * reward
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
total_loss += loss.item()
return total_loss / num_steps
def train_hierarchical_planning(
self,
tasks: List[Dict],
num_steps: int = 1000
) -> float:
"""
Train skill: decompose multi-step questions into tool sequence.
Task: Given question and facts, predict tool call order.
"""
total_loss = 0.0
for step in range(num_steps):
task = random.choice(tasks)
question = task['question']
tool_sequence = task['tool_sequence']
# Predict full tool sequence
planning_prompt = f"Question: {question}\n\nPlan the search steps:"
predicted_plan, log_probs = self.model.generate_with_logprobs(
planning_prompt, max_tokens=200
)
# Reward: how many tools in correct order?
num_correct = self._score_plan(predicted_plan, tool_sequence)
reward = num_correct / max(1, len(tool_sequence))
# Update
loss = -log_probs.mean() * reward
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
total_loss += loss.item()
return total_loss / num_steps
def _match_grounding(self, prediction: str, expected_tool: str) -> bool:
"""Check if prediction matches expected tool."""
return expected_tool.lower() in prediction.lower()
def _score_plan(self, prediction: str, expected_sequence: List[Dict]) -> int:
"""Count correctly ordered tools in prediction."""
correct = 0
for tool_dict in expected_sequence:
if tool_dict['tool'] in prediction:
correct += 1
return correct
Step 4: Environment interaction stage
class LongHorizonSearchRL:
"""RL training with actual environment interaction."""
def __init__(
self,
model,
optimizer,
environment,
max_steps: int = 30
):
self.model = model
self.optimizer = optimizer
self.env = environment
self.max_steps = max_steps
def generate_trajectory(
self,
task: Dict,
temperature: float = 0.7
) -> Dict:
"""
Generate trajectory by interacting with environment.
Returns:
trajectory: {question, steps, observations, reward, success}
"""
question = task['question']
trajectory = {
'question': question,
'steps': [],
'observations': [],
'success': False
}
current_state = question
discovered_facts = set()
for step in range(self.max_steps):
# Predict next action
prompt = f"Question: {question}\nCurrent findings: {current_state}\n\nNext action:"
action, log_probs = self.model.generate_with_logprobs(
prompt, max_tokens=50, temperature=temperature
)
trajectory['steps'].append({
'action': action,
'log_probs': log_probs
})
# Execute in environment
observation = self.env.execute_action(action)
trajectory['observations'].append(observation)
# Check for success
current_state = f"{current_state}\n{observation}"
for fact in task['gold_facts']:
if fact in observation:
discovered_facts.add(fact)
if len(discovered_facts) == len(task['gold_facts']):
trajectory['success'] = True
break
trajectory['num_facts_found'] = len(discovered_facts)
trajectory['reward'] = 1.0 if trajectory['success'] else 0.5 * (len(discovered_facts) / len(task['gold_facts']))
return trajectory
def update_from_trajectory(self, trajectory: Dict):
"""Update policy from trajectory."""
total_loss = 0.0
for step in trajectory['steps']:
log_probs = step['log_probs']
# Discounted reward
reward = trajectory['reward']
loss = -log_probs.mean() * reward
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
total_loss += loss.item()
return total_loss / len(trajectory['steps'])
Step 5: Staged training pipeline
def train_redsearcher_staged(
model,
corpus: List[str],
entity_graph,
optimizer,
num_task_synthesis_iter: int = 500,
num_skill_steps: int = 5000,
num_env_steps: int = 10000,
device: str = 'cuda'
):
"""
Full REDSearcher training: synthesis → skills → environment.
"""
# Stage 1: Task synthesis
print("Stage 1: Task Synthesis")
synthesizer = DualConstraintedTaskSynthesis(corpus, entity_graph)
tool_converter = ToolAugmentedConstraintRepresentation({})
tasks = []
for i in range(num_task_synthesis_iter):
# Vary complexity
treewidth = (i % 5) + 1
dispersion = 0.5 + 0.3 * (i % 10) / 10
task = synthesizer.synthesize_query_with_complexity(
target_treewidth=treewidth,
information_dispersion=dispersion
)
task['tool_sequence'] = tool_converter.create_tool_sequence(task)
tasks.append(task)
print(f" Generated {len(tasks)} tasks")
# Stage 2: Atomic skill learning
print("Stage 2: Atomic Skill Training")
skill_trainer = AtomicSkillTrainer(model, optimizer, {})
loss_grounding = skill_trainer.train_intent_grounding(tasks, num_skill_steps // 2)
loss_planning = skill_trainer.train_hierarchical_planning(tasks, num_skill_steps // 2)
print(f" Grounding loss: {loss_grounding:.4f}")
print(f" Planning loss: {loss_planning:.4f}")
# Stage 3: Environment interaction
print("Stage 3: Environment Interaction")
env_trainer = LongHorizonSearchRL(model, optimizer, {}, max_steps=30)
total_success = 0
for step in range(num_env_steps):
task = random.choice(tasks)
trajectory = env_trainer.generate_trajectory(task)
loss = env_trainer.update_from_trajectory(trajectory)
if trajectory['success']:
total_success += 1
if (step + 1) % 1000 == 0:
success_rate = total_success / (step + 1)
print(f" Step {step + 1}: Success rate={success_rate:.2%}")
return model
Practical Guidance
When to use: Multi-step information-seeking agents; questions requiring 10+ search steps; domains with clear document retrieval
Hyperparameters:
- treewidth_range: 1-5 (complexity)
- information_dispersion: 0.4-0.9 (sparsity)
- max_steps: 20-40 (trajectory length)
- skill_training_ratio: Allocate steps 40% grounding, 40% planning, 20% integration
Key advantages:
- Systematic curriculum from simple to complex
- Staged training reduces expensive environment calls
- Atomic skills enable transfer across domains
- Long-horizon support (30+ steps)
Common pitfalls:
- Too much task complexity too soon → exploration failure
- Skill training too brief → weak foundations
- Not validating facts are actually retrievable
- Environment cost not amortized enough
Scaling: Simulation environment enables large-scale training; real interaction only for final polish.
Reference
Paper: https://arxiv.org/abs/2602.14234 Related work: Information retrieval, long-horizon planning, RL for agents Benchmarks: Humanity's Last Exam, BrowseComp, custom search tasks