• 学术相关
  • Intern-Atlas
    • 0. Execution Overview
    • 1. High-level Design (Indexing → Retrieval → Generation)
    • 2. Offline Construction: Indexing (Detailed Execution)
    • 3. Online Query: Retrieval (Detailed Execution)
    • 4. Online Generation: Generation (Detailed Execution)
    • 5. Key Design Decisions
    • 6. Evaluation
    • 7. Limitations and Applicability
    • 8. Quick Reference
  • Search-R1
    • 0. Execution Overview
    • 1. High-level Design (Indexing → Retrieval → Generation)
    • 2. Offline Construction: RL Training (Detailed Execution)
    • 3. Online Query: Retrieval (Detailed Execution)
    • 4. Online Generation: Generation (Detailed Execution)
    • 5. Key Design Decisions
    • 6. Evaluation
    • 7. Limitations and Applicability
    • 8. Quick Reference

学术相关

Intern-Atlas

One-sentence positioning: Extracts “method entities + typed causal edges + bottleneck/mechanism evidence” from 1M+ AI papers to construct a queryable methodological evolution graph, serving as the underlying knowledge infrastructure for AI research agents.

Key innovation: Upgrades flat citation networks into a “method-method causal graph” where each causal edge is accompanied by verbatim citations and structured bottleneck/mechanism annotations; proposes the SGT-MCTS algorithm to reconstruct method evolution lineages on this graph, enabling idea evaluation and generation based on explicit structural evidence rather than LLM parametric memory.

0. Execution Overview

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Offline Phase (graph construction)
  ├─ ① Paper corpus processing (1,030,314 papers)
  ├─ ② Method entity extraction + alias resolution
  │     ├─ Seed method curation
  │     ├─ LLM Proposer scanning expansion
  │     └─ Alias registry A: 8,155 canonical / 9,545 aliases
  ├─ ③ Citation edge semantic classification (7 types)
  │     ├─ strong-causal: extends / improves / replaces / adapts
  │     ├─ non-strong: uses_component / compares
  │     └─ non-causal: background
  └─ ④ Causal edge evidence filling
        ├─ 14-category bottleneck taxonomy b_e
        ├─ Mechanism m_e / trade-off t_e
        ├─ LLM confidence c_e ∈ [0,1]
        └─ Verbatim citation grounding
              ↓
Online Phase (retrieval + generation)
  ├─ ⑤ Node matching (keywords + BM25)
  ├─ ⑥ Lineage Reconstruction: SGT-MCTS
  │     ├─ Forward/backward dual trees
  │     ├─ SGT-UCT selection = standard UCT + λ·graph-aware prior
  │     ├─ Temporal coherence TC + edge confidence conf joint prior
  │     └─ Branch discovery + Jaccard deduplication
  ├─ ⑦ Lineage Rerank (length + evidence strength + search consensus)
  └─ ⑧ Generation layer
        ├─ Graph-Grounded Idea Evaluator (5-dimension scoring)
        └─ Graph-Grounded Idea Generator (4 strategy types + evidence certificate)

1. High-level Design (Indexing → Retrieval → Generation)

1.1 Indexing

1.2 Retrieval

1.3 Generation

2. Offline Construction: Indexing (Detailed Execution)

Step 2.1 Corpus Preparation

Step 2.2 Method Entity Extraction and Alias Resolution

Step 2.2.1 Seed Method Construction

Step 2.2.2 Method Expansion

Step 2.2.3 Alias Registry Construction

Step 2.3 Citation Edge Semantic Classification (Two-Phase LLM Extraction)

Step 2.3.1 Phase 1: Edge Type Classification

Step 2.3.2 Phase 2: Structured Record Completion

Step 2.4 Causal Edge Evidence Filling

Step 2.4.1 14-Category Bottleneck Taxonomy

Step 2.4.2 Structured Evidence Quadruple

Each causal edge e carries quadruple ρ(e) = (b_e, m_e, t_e, c_e):

Step 2.4.3 Verbatim Citation Verification

Step 2.4.4 Graph Scale

3. Online Query: Retrieval (Detailed Execution)

3.1 Retrieval Mode Overview

Intern-Atlas employs a single lineage reconstruction retrieval mode, with SGT-MCTS searching on the method evolution graph:

3.2 Retrieval Procedure

Step 3.1: Node Matching

Step 3.2: SGT-MCTS Lineage Reconstruction

Step 3.2.1: Dual-tree Construction

Step 3.2.2: SGT-UCT Selection

Calibration range limitation: TC is only calibrated on post-2015 AI literature; domains with different research paces require recalibration.

Step 3.2.3: Expansion (Confidence-prioritized)

Step 3.2.4: Rollout (Greedy)

Step 3.2.5: Backpropagation

Step 3.2.6: Path Stitching and Deduplication

Step 3.2.7: Branch Discovery

Step 3.3: Lineage Rerank

3.3 Retrieval Algorithm Comparison

4. Online Generation: Generation (Detailed Execution)

4.1 Generation Procedure

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Input: query q, retrieved evolution paths Π_q
    │
    ▼
┌─────────────────────────────┐
│ Parse methods mentioned     │
│ in idea d, map to nodes M_d │
│ via lookup table A          │
└─────────────┬───────────────┘
              │
         ┌────┴────┐
         ▼         ▼
   Evaluator   Generator
   (Evaluate)   (Generate)
      │           │
      ▼           ▼
  5-dim scoring  4 strategy types
  +cross penalty +evidence certificate

Step 4.1: Graph-Grounded Idea Evaluator

Step 4.1.1: Motivation

Free-text LLM judges prefer stacking popular methods; novelty scores are negatively correlated with actual scientific impact; graph statistics can directly provide structural evidence.

Step 4.1.2: Per-dimension Scoring

Formula:

Step 4.1.3: Cross-dimensional Aggregation

Formula:

Step 4.1.4: Idea Evaluation Results

Step 4.2: Graph-Grounded Idea Generator

Step 4.2.1: Structural Gap Pattern → Generation Strategy Mapping

Step 4.2.2: Evidence Certificate (Core Anti-hallucination Mechanism)

5. Key Design Decisions

6. Evaluation

6.1 Graph Construction Quality

Benchmark: 30 high-impact surveys → 30 method evolution graphs, containing 2,268 nodes / 1,462 edges / 133 evolution chains.

6.2 Idea Evaluator (Strata Dataset)

Evaluation method: Publication tier verification (across publication strata) + human evaluation.

6.3 Idea Generator Comparative Experiments

6.4 SGT-MCTS Retrieval Algorithm Comparison

7. Limitations and Applicability

Best Applicable Scenarios

• AI agents performing idea generation/evaluation that require structured method evolution priors
• Traceable scientific idea generation (evidence certificate prevents hallucination)
• Method lineage research / survey automation within AI subfields
• Evaluating ideas while avoiding the “popular method stacking” bias of LLM free-text judges

Unsuitable Scenarios

• Disciplines with research paces significantly different from AI (TC calibration fails)
• Extremely niche subfields without human survey references (limited graph coverage)
• Research focused on paper-level citation network properties (e.g., PageRank, influence propagation)

8. Quick Reference

Search-R1

One-sentence positioning: A framework that trains LLMs via reinforcement learning to autonomously generate multi-turn search queries during step-by-step reasoning and perform real-time retrieval, addressing the suboptimal performance of prompt-engineering-driven search through retrieval result loss masking and outcome-oriented rewards.

Key innovation: Unlike prompt-driven search methods, Search-R1 models the search engine as part of the environment and trains the LLM via RL to autonomously learn the optimal search interaction strategy; simultaneously introduces retrieval token masking to prevent gradient updates on retrieved content, ensuring training stability.

0. Execution Overview

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Training Phase
  ├─ ① Preparation: search engine as environment, policy model π_θ, reference model π_ref
  ├─ ② Rollout: model generates reasoning sequences, interleaving token generation with search engine retrieval
  │   ├─ Encountering uncertain knowledge → generate <search>... </search> query
  │   ├─ Search engine returns documents → wrap as <information>... </information>
  │   ├─ Reasoning process → wrap as <think>... </think>
  │   └─ Final answer → wrap as <answer>... </answer>
  ├─ ③ Reward computation: outcome-oriented reward r_φ(x, y) = EM(a_pred, a_gold)
  ├─ ④ Loss Masking: policy gradient computed only on LLM-generated tokens; retrieved content excluded from optimization
  └─ ⑤ Policy update: PPO or GRPO updates the policy model

Inference Phase (per query)
  ├─ ① Input question q
  ├─ ② Iterative generation: text generation ↔ search query alternation
  │   ├─ Generate response tokens
  │   ├─ Detect <search> / </answer> / <eos> → determine next action
  │   ├─ Search query → retrieve documents → inject into reasoning chain
  │   └─ Continue generation
  ├─ ③ Termination condition: maximum action count B reached or <answer> generated
  └─ ④ Output final answer

1. High-level Design (Indexing → Retrieval → Generation)

1.1 Indexing

1.2 Retrieval

1.3 Generation

2. Offline Construction: RL Training (Detailed Execution)

The core of Search-R1’s offline phase is RL training rather than traditional knowledge base index construction. The system relies on external search engines (e.g., Bing) for document retrieval, and the notes do not explicitly record traditional index construction steps.

Step 2.1 Search Engine Environment Modeling

Step 2.2 Structured Token System Definition

Step 2.3 Rollout Sequence Generation

Step 2.4 Retrieval Token Masking (Loss Masking)

Step 2.5 Reward Computation

Step 2.6 Policy Update (PPO / GRPO)

3. Online Query: Retrieval (Detailed Execution)

3.1 Retrieval Mode Overview

Search-R1 employs a single RL-driven search mode, autonomously triggered by the trained model during inference:

3.2 Retrieval Procedure

Step 3.1: Reasoning and Search Alternation

Step 3.2: Search Query Detection and Execution

Step 3.3: Iterative Termination Judgment

4. Online Generation: Generation (Detailed Execution)

4.1 Reasoning Procedure

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Input question q
    │
    ▼
┌─────────────────────────┐
│ Initialize: action count b ← 0 │
│ Initialize: response y ← ∅      │
└───────────┬─────────────┘
            │
            ▼
    ┌───────┴───────┐
    │  while b < B  │
    └───────┬───────┘
            │
            ▼
    ┌───────────────────┐
    │ Generate response tokens │
    │ until termination signal │
    └─────────┬─────────┘
              │
         ┌────┴────┐
         ▼         ▼
      <search>   <answer>
         │         │
         ▼         ▼
      Retrieve    Return y
      documents
      inject
         │
         ▼
      b ← b + 1
         │
         └──────→ Continue generating

Step 4.1: Online Reasoning Sequence Generation

5. Key Design Decisions

6. Evaluation

6.1 Evaluation Metrics

6.2 RL Training Comparison

6.3 Experimental Setup

6.4 Datasets

7. Limitations and Applicability

The notes do not explicitly record Search-R1’s limitations.

Best Applicable Scenarios

• Reasoning tasks requiring efficient acquisition of external knowledge and up-to-date information
• Open-domain QA (NQ, TriviaQA, etc.)
• Multi-hop QA (HotpotQA, 2WikiMultiHopQA, MuSiQue, etc.)
• Knowledge-intensive reasoning tasks

Unsuitable Scenarios

• —

8. Quick Reference

• Previous
  【阅】本周阅读摘选2026-05-18 → 2026-05-24