Skip to main content الرئيسية المنشئون adu2021 skillxiv recurrence-memory-reasoning-depth
recurrence-memory-reasoning-depth Extend neural network reasoning capabilities through recurrence (repeated computation cycles), external memory (intermediate state storage), and test-time compute scaling for multi-step reasoning.
الانتقال إلى التثبيت سوق المهارات اكتشف واستكشف مهارات الذكاء الاصطناعي التي بناها المجتمع.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
عرض تفاصيل Prompt يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/ADu2021/skillXiv --skill recurrence-memory-reasoning-depthيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
تحميل Zip جاري التحميل... المزيد من هذا المستودع meaningful-kebab-case-name Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
action-quantization-behavior-cloning Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
adaptive-lora-personalized-ranks Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
المهن ذات الصلة SOC
استنادا إلى تصنيف SOC المهني
name recurrence-memory-reasoning-depth title Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2508.16745 keywords ["reasoning-depth","recurrent-computation","external-memory","test-time-scaling","sequential-reasoning"] description Extend neural network reasoning capabilities through recurrence (repeated computation cycles), external memory (intermediate state storage), and test-time compute scaling for multi-step reasoning.
Extending Reasoning Depth with Recurrence, Memory, and Compute Scaling
Core Concept
Multi-step reasoning requires computational depth beyond what fixed-depth architectures provide. This skill combines three mechanisms: recurrence (allowing repeated computation passes), external memory (storing intermediate states), and test-time compute scaling (allocating more cycles during inference) to extend effective reasoning depth. Studies on cellular automata and Boolean functions show that while models can memorize next-step predictions, multi-step reasoning requires explicit recurrent computation.
Architecture Overview
Recurrent Processing : Iterative computation allowing state evolutionExternal Memory System : Storage for intermediate reasoning statesTest-Time Compute Allocation : Dynamic cycle budget during inferenceDepth Extension : Exceeding architectural layer countSequential Rule Learning : Multi-step state transition modeling
Implementation Steps
1. Implement Recurrent Computation Unit
Create iterative computation mechanism:
import torch
import torch.nn as nn
from typing import Tuple , List , Dict
class RecurrentReasoningUnit (nn.Module):
"""Iterative computation for multi-step reasoning."""
def __init__ (
self,
state_dim: int = 256 ,
hidden_dim: int = 512 ,
num_recurrence_steps: int = 5
):
super ().__init__()
self .state_dim = state_dim
self .num_recurrence_steps = num_recurrence_steps
self .recurrent_cell = nn.GRUCell(state_dim, hidden_dim)
self
self
self
self
def
forward
self,
initial_state: torch.Tensor,
num_steps: int = None ,
external_memory: torch.Tensor = None
Tuple
List
"""
Iterate computation for multiple recurrence steps.
"""
if
is
None
self
0
self
for
in
range
self
self
self
self
self
1
return
def
compute_reasoning_depth
self,
num_steps: int
float
"""
Compute effective reasoning depth.
Recurrent models achieve depth > number of layers.
"""
1
return
2. Implement External Memory System Store and retrieve intermediate states:
class ExternalMemory (nn.Module):
"""External memory for storing reasoning states."""
def __init__ (
self,
state_dim: int = 256 ,
memory_size: int = 32 ,
num_read_heads: int = 4
):
super ().__init__()
self .state_dim = state_dim
self .memory_size = memory_size
self .num_read_heads = num_read_heads
self .register_buffer("memory" , torch.randn(memory_size, state_dim) * 0.01 )
self .write_controller = nn.Linear(state_dim, memory_size)
self .read_query = nn.Linear(state_dim, state_dim)
self .key_proj = nn.Linear(state_dim, state_dim)
def write_to_memory (
self,
state: torch.Tensor,
write_addr: torch.Tensor = None
) -> torch.Tensor:
"""
Write state to memory with soft addressing.
"""
batch_size = state.shape[0 ]
if write_addr is None :
write_logits = self .write_controller(state)
write_weights = torch.softmax(write_logits, dim=-1 )
else :
write_weights = write_addr
for i in range (batch_size):
for j in range (self .memory_size):
self .memory[j] = 0.9 * self .memory[j] + 0.1 * write_weights[i, j] * state[i]
return write_weights
def read_from_memory (
self,
query: torch.Tensor,
num_read_heads: int = None
) -> torch.Tensor:
"""
Read from memory using content-based addressing.
"""
if num_read_heads is None :
num_read_heads = self .num_read_heads
batch_size = query.shape[0 ]
query_proj = self .read_query(query)
key_proj = self .key_proj(self .memory)
read_output = torch.zeros(batch_size, self .state_dim)
for head in range (num_read_heads):
head_query = query_proj[:, :self .state_dim // num_read_heads]
head_keys = key_proj[:, :self .state_dim // num_read_heads]
scores = torch.matmul(head_query, head_keys.t())
weights = torch.softmax(scores / (self .state_dim ** 0.5 ), dim=-1 )
head_read = torch.matmul(weights, self .memory)
read_output[:, head * self .state_dim // num_read_heads:(head + 1 ) * self .state_dim // num_read_heads] = head_read[:, :self .state_dim // num_read_heads]
return read_output
def reset_memory (self ):
"""Clear memory for new reasoning episode."""
self .memory.zero_()
3. Implement Test-Time Compute Scaling Allocate variable computation budget during inference:
class TestTimeComputeScaler :
"""Dynamically allocate compute budget at test time."""
def __init__ (
self,
base_steps: int = 5 ,
max_steps: int = 50 ,
budget_per_example: float = 1.0
):
self .base_steps = base_steps
self .max_steps = max_steps
self .budget_per_example = budget_per_example
def get_step_budget (
self,
task_description: str ,
task_complexity: float = 0.5
) -> int :
"""
Compute number of recurrence steps based on task complexity.
"""
complexity_budget = int (self .base_steps * (1.0 + task_complexity * 5.0 ))
budget = min (complexity_budget, self .max_steps)
return budget
def adaptive_compute (
self,
model: "ReasoningModel" ,
state: torch.Tensor,
task: str ,
target_accuracy: float = 0.95 ,
max_total_steps: int = 50
) -> Tuple [torch.Tensor, Dict [str , any ]]:
"""
Iteratively increase compute until convergence or budget exhausted.
"""
current_state = state
step = 0
previous_output = None
convergence_reached = False
metrics = {"total_steps" : 0 , "convergence_step" : None , "final_output_stable" : False }
while step < max_total_steps and not convergence_reached:
current_state, trajectory = model.recurrence_unit(current_state, num_steps=1 )
current_output = model.output_head(current_state)
if previous_output is not None :
similarity = torch.nn.functional.cosine_similarity(
current_output.unsqueeze(0 ),
previous_output.unsqueeze(0 )
).item()
if similarity > target_accuracy:
convergence_reached = True
metrics["convergence_step" ] = step
metrics["final_output_stable" ] = True
previous_output = current_output
step += 1
metrics["total_steps" ] = step
return current_state, metrics
def measure_reasoning_quality (
self,
trajectories: List [List [torch.Tensor]],
ground_truth: torch.Tensor
) -> Dict [str , float ]:
"""
Measure quality of multi-step reasoning.
"""
metrics = {
"next_step_accuracy" : 0.0 ,
"multistep_accuracy" : 0.0 ,
"trajectory_stability" : 0.0
}
for trajectory in trajectories:
for t in range (len (trajectory) - 1 ):
pred_next = trajectory[t + 1 ]
actual_next = ground_truth[t + 1 ] if t + 1 < len (ground_truth) else None
if actual_next is not None :
match = torch.allclose(pred_next, actual_next, atol=1e-5 )
metrics["next_step_accuracy" ] += match
final_pred = trajectory[-1 ]
final_actual = ground_truth[-1 ] if len (ground_truth) > 0 else None
if final_actual is not None :
match = torch.allclose(final_pred, final_actual, atol=1e-5 )
metrics["multistep_accuracy" ] += match
stacked = torch.stack(trajectory)
variance = torch.var(stacked, dim=0 ).mean().item()
stability = 1.0 / (1.0 + variance)
metrics["trajectory_stability" ] += stability
n = len (trajectories)
for key in metrics:
metrics[key] = metrics[key] / n if n > 0 else 0.0
return metrics
4. Integrate Recurrence, Memory, and Compute class DeepReasoningModel (nn.Module):
"""Complete model with recurrence, memory, and test-time scaling."""
def __init__ (
self,
input_dim: int = 256 ,
state_dim: int = 256 ,
hidden_dim: int = 512 ,
memory_size: int = 32
):
super ().__init__()
self .input_proj = nn.Linear(input_dim, state_dim)
self .recurrence_unit = RecurrentReasoningUnit(state_dim, hidden_dim)
self .memory = ExternalMemory(state_dim, memory_size)
self .output_head = nn.Linear(state_dim, input_dim)
self .compute_scaler = TestTimeComputeScaler()
def forward (
self,
input_state: torch.Tensor,
num_steps: int = 5 ,
use_memory: bool = True
) -> torch.Tensor:
"""Forward pass with recurrence and memory."""
state = self .input_proj(input_state)
for step in range (num_steps):
state, _ = self .recurrence_unit(state, num_steps=1 )
if use_memory:
self .memory.write_to_memory(state)
if use_memory:
memory_context = self .memory.read_from_memory(state)
state = state + 0.3 * memory_context
output = self .output_head(state)
return output
def solve_with_adaptive_compute (
self,
input_state: torch.Tensor,
task_description: str
) -> Tuple [torch.Tensor, Dict ]:
"""Solve with test-time compute scaling."""
complexity = self ._estimate_complexity(task_description)
budget = self .compute_scaler.get_step_budget(task_description, complexity)
output, metrics = self .compute_scaler.adaptive_compute(
self ,
input_state,
task_description,
max_total_steps=budget
)
return output, metrics
def _estimate_complexity (self, task: str ) -> float :
"""Estimate task complexity from description."""
complexity_indicators = ["recursive" , "nested" , "multi-step" , "chain" ]
complexity = sum (1 for ind in complexity_indicators if ind in task.lower()) / len (complexity_indicators)
return min (1.0 , complexity)
Practical Guidance
When to Use Deep Reasoning Models
Multi-step mathematical reasoning
Algorithmic problem solving
Abstract rule learning
Tasks requiring state accumulation
Scenarios allowing test-time compute allocation
When NOT to Use
Simple single-step generation
Real-time systems with strict latency (<100ms)
Tasks without clear sequential structure
Inference with extremely limited budgets
Key Hyperparameters
num_recurrence_steps : 5-20 (base steps)memory_size : 16-64 slotsstate_dim : 128-512max_test_time_steps : 20-100 (dependent on budget)convergence_threshold : 0.90-0.99
Performance Expectations
Next-Step Accuracy: Maintained with recurrence
Multi-Step Accuracy: Significantly improved
Effective Depth: Layer count × recurrence steps
Memory Efficiency: Sub-quadratic vs. attention
Reference Researchers. (2024). Beyond Memorization: Extending Reasoning Depth with Recurrence Memory and Test-Time Compute Scaling. arXiv preprint arXiv:2508.16745.
