概念
retrieval reasoning agent inference benchmark
创建2026-05-05
更新2026-05-05
阅读量级16 分钟
概念导读

Large reasoning models (LRMs) like DeepSeek-R1 and OpenAI o1 generate extended chains of thought spanning 12,000-25,000 tokens, yet their integration with retrieval-augmented generation (RAG) is fundamentally misaligned.

  1. Core Problem Definition
  2. Method: ReaLM-Retrieve Three Components
  3. Component 1: Step-Level Uncertainty Detection (RSUS)
  4. Component 2: Reasoning Step Segmentation
  5. Component 3: Learned Retrieval Intervention Policy
  6. Efficient Retrieval Integration
  7. Experimental Results
  8. Main Results (Multi-hop QA)

ReaLM-Retrieve: When to Retrieve During Reasoning

Core Problem Definition

Large reasoning models (LRMs) like DeepSeek-R1 and OpenAI o1 generate extended chains of thought spanning 12,000-25,000 tokens, yet their integration with retrieval-augmented generation (RAG) is fundamentally misaligned.

The temporal mismatch: Current RAG pipelines optimize for providing context before reasoning begins ("retrieve-then-generate"). But reasoning models encounter knowledge gaps during multi-step inference chains — gaps that cannot be anticipated at query time.

Three incompatibilities with existing iterative retrieval: 1. Granularity mismatch: Token-level (FLARE) or sentence-level (IRCoT) vs. reasoning-step granularity 2. Signal unavailability: Token probability methods (FLARE) require logits unavailable from completion-only models (o1) 3. Efficiency collapse: Fixed-interval retrieval (IRCoT) adds 2-5s per call, prohibitive for 12K-25K token chains

Formal Statement: Given query q, document collection D, reasoning model generating chain R = (r_1, ..., r_n) where each r_i is a reasoning step (logical unit spanning multiple sentences), learn an intervention policy π that at each step decides: - Retrieve decision: a^ret_i ∈ {0,1} - Query formulation: q_i = f(q, r_{1:i}) - Context integration: c_i = g(r_{1:i}, R_i)

Objective: max_π E_{q~Q}[Acc(a_π, a*) - λ · Cost(π)]

Method: ReaLM-Retrieve Three Components

Component 1: Step-Level Uncertainty Detection (RSUS)

Reasoning Step Uncertainty Score — a composite of three signals:

RSUS(r_i) = α · U_verb(r_i) + β · U_ent(r_i) + γ · U_cons(r_i)

with learned weights (α, β, γ) = (0.40, 0.35, 0.25) maximizing correlation (r=0.72) with downstream retrieval benefit.

U_verb — Verbalized Uncertainty: Prompt at step boundaries: "Rate your confidence that the current conclusion is factually correct (0-100)". Normalized to [0,1]. For completion-only models: 2-layer MLP proxy trained on SBERT embedding + surface features (hedge phrases, named entities), achieving 0.71 AUROC vs 0.83 for true verbalized confidence.

U_ent — Entity-based Entropy: Extract named entities from r_i, compute retrieval score entropy across candidate documents:

U_ent(r_i) = -Σ_{e∈Ent(r_i)} p(e|D) log p(e|D)

High entropy = entities with ambiguous/sparse corpus coverage. Uses BM25 scores from top-100 documents per entity.

U_cons — Consistency Signal: Sample k=3 alternative continuations at critical steps (identified by discourse markers), measure agreement. Disagreement indicates reasoning uncertainty.

Overhead: RSUS adds only 8% to reasoning inference vs 500-2000% for semantic entropy.

Component 2: Reasoning Step Segmentation

Lightweight classifier (3-layer transformer encoder, hidden dim 256, 4 attention heads) operating on 128-token sliding windows with 64-token stride. Predicts boundary probability from: 1. Discourse markers ("therefore", "however", "this means") 2. Logical connectives indicating inference completion 3. Topic shifts via embedding similarity 4. Punctuation and formatting patterns

Training: 2,847 human-annotated reasoning traces from DeepSeek-R1 on NaturalQuestions. Inter-annotator agreement κ=0.78. Trained 10 epochs, lr=5×10^{-5}, batch size 32. Achieves 94.2% F1 on held-out test (N=412 traces). Average segment: 127 tokens vs 23 for sentence-level.

Generalization: On out-of-domain benchmarks, boundary F1 degrades modestly (1.5-3.0 points): MuSiQue 91.4%, HotpotQA 92.7%, 2WikiMultiHopQA 91.2%.

Component 3: Learned Retrieval Intervention Policy

Modeled as contextual bandit with state s_i = (q, r_{1:i}, RSUS(r_i), h_i) where h_i encodes retrieval history.

a^ret_i = 1[π_θ(s_i) > τ]    (τ = 0.65)
q_i = QueryGen(s_i) if a^ret_i = 1

Policy architecture: Lightweight transformer encoder processing concatenation of query embedding, current reasoning step embedding, RSUS features, and retrieval history. Outputs retrieval probability + query representation.

Training: REINFORCE with reward R = F1(a_π, a*) - λ_1 · n_ret - λ_2 · t_latency. QueryGen receives gradients via same reward signal. Curriculum learning: start with high λ (penalize retrieval), gradually decrease.

QueryGen: Single-layer transformer decoder (hidden dim 512, 8 heads) cross-attending to current reasoning step. Queries formulated as information needs ("What is [entity]'s relationship to [concept]?") outperform direct reasoning-step queries.

Efficient Retrieval Integration

Implicit compression: Attention-weighted importance scores for retrieved passages, retain only sentences exceeding threshold τ_rel=0.45. Reduces context expansion by 73% while preserving 96% of retrieval utility.

Speculative caching: After first retrieval, speculatively retrieve documents for likely next-step queries based on entities in retrieved content. 37% hit rate eliminates retrieval latency entirely for those calls.

KV-cache preservation: For open-weight models, preserves cached key-value states for unchanged context portions. Reduces time-to-first-token after retrieval by 2.1x.

Combined effect: 3.2x per-call efficiency improvement (0.66s vs 2.10s naive), enabling 1.33x lower end-to-end latency than IRCoT.

Experimental Results

Main Results (Multi-hop QA)

DeepSeek-R1-Distill-Qwen-32B:

Method MuSiQue F1 MuSiQue Calls HotpotQA F1 HotpotQA Calls 2WikiMHQA F1 2WikiMHQA Calls Avg F1
No Retrieval 48.7 0.0 51.2 0.0 47.3 0.0 49.1
Single RAG 59.4 1.0 62.8 1.0 60.1 1.0 60.8
IRCoT 65.4 3.4 67.9 4.1 66.2 3.8 66.5
FLARE* 62.3 2.8 65.4 3.2 63.7 2.9 63.8
Self-RAG† 61.9 2.1 66.3 2.4 64.5 2.2 64.2
Search-R1 66.8 2.4 69.2 2.7 67.4 2.5 67.8
ReaLM-Retrieve 71.2 1.8 71.8 1.9 69.7 1.7 70.9

All ReaLM-Retrieve improvements significant at p<0.01 (paired bootstrap, 10K iterations, Bonferroni corrected).

DeepSeek-R1-671B:

Method MuSiQue F1 Calls HotpotQA F1 Calls Avg F1
Search-R1 73.4 2.6 75.2 2.8 74.1
ReaLM-Retrieve 77.8 1.9 78.4 2.0 77.6

4.4% absolute improvement over Search-R1 (p<0.01).

Statistical Significance (R1-32B)

Dataset ΔF1 vs IRCoT 95% CI ΔF1 vs Search-R1 95% CI
MuSiQue +5.8 [4.2, 7.4] +4.4 [2.9, 5.9]
HotpotQA +3.9 [2.4, 5.4] +2.6 [1.2, 4.0]
2WikiMHQA +3.5 [2.1, 4.9] +2.3 [0.9, 3.7]

Efficiency Analysis (MuSiQue, R1-32B)

Method Calls Latency (s) Per-Call (s) Tokens F1/Call
No Retrieval 0.0 12.4 8,432
Single RAG 1.0 13.2 0.80 9,156 59.4
Naive Interleave 4.2 21.2 2.10 12,847 14.9
IRCoT 3.4 18.7 1.85 11,284 19.2
FLARE 2.8 16.9 1.61 10,647 22.3
Search-R1 2.4 15.8 1.42 10,102 27.8
ReaLM-Retrieve 1.8 14.1 0.66 9,489 39.6

Cost savings vs IRCoT: 16% (conservative, output-token only at $2.19/1M tokens).

Ablation Study (MuSiQue, R1-32B)

Configuration F1 Calls ΔF1
Full ReaLM-Retrieve 71.2±0.6 1.8±0.1
Uncertainty ablation
w/o U_verb 68.4±0.8 2.1±0.2 -2.8**
w/o U_ent 69.1±0.7 1.9±0.1 -2.1**
w/o U_cons 70.3±0.5 1.8±0.1 -0.9*
RSUS → Random 62.7±1.2 1.8±0.1 -8.5**
Policy ablation
Fixed threshold τ 68.9±0.9 2.4±0.2 -2.3**
w/o Query formulation 69.4±0.6 1.8±0.1 -1.8**
w/o Retrieval history 69.8±0.7 2.0±0.1 -1.4*
Integration ablation
w/o Implicit compression 70.8±0.5 1.8±0.1 -0.4
w/o Speculative cache 71.0±0.6 1.8±0.1 -0.2
w/o KV preservation 70.9±0.5 1.8±0.1 -0.3

Key finding: Verbalized uncertainty contributes most (-2.8% F1 when removed). Random triggering degrades by 8.5%, proving when to retrieve matters enormously.

Performance by Question Complexity

Hops ReaLM-Retrieve F1 IRCoT F1 ΔF1
2-hop 72.1 68.9 +3.2
3-hop 69.4 63.6 +5.8
4-hop 66.2 57.8 +8.4

Advantage increases with complexity: step-level uncertainty detection provides most value on longer reasoning chains.

Training/Implementation Configuration

  • Hardware: 8× NVIDIA A100 80GB GPUs
  • Reasoning inference: vLLM with tensor parallelism
  • Retrieval: ColBERTv2 with PLAID engine, WARP optimizations (171ms avg latency)
  • Policy training: 50,000 steps, lr=10^{-4}, batch size 64, curriculum-decreasing λ_1 from 0.5 to 0.1
  • RSUS weights: (α, β, γ) = (0.40, 0.35, 0.25) via grid search on validation split
  • Compression threshold: τ_rel = 0.45 (retains 27% of content, preserves 96% utility)
  • Retrieval threshold: τ = 0.65 (balances precision 0.72 / recall 0.81)
  • Statistical testing: Paired bootstrap (10K iterations), Bonferroni correction, 3 seeds (42, 123, 456)
Method Granularity Black-box Compatible RL Training Step-Level Uncertainty Efficiency Optimized
Single RAG Query N/A
IRCoT Sentence
FLARE Token ❌ (needs logits)
DRAGIN Token ❌ (needs attention)
Self-RAG Sentence ❌ (needs fine-tuning) Partial
Search-R1 Step Partial
ReaLM-Retrieve Reasoning step ✅ RSUS ✅ 3.2x

Critical Analysis

Strengths

  1. Principled problem formulation: Clearly identifies the temporal mismatch between RAG and reasoning models
  2. RSUS is elegant: Three complementary uncertainty signals (verbalized, entity-entropy, consistency) with only 8% overhead vs 500-2000% for semantic entropy
  3. Strong empirical results: +10.1% F1 with 47% fewer retrieval calls — establishes new Pareto frontier
  4. Comprehensive ablation: Each component's contribution isolated, random triggering degrades 8.5% — proves the method works
  5. Scales with model capability: Results improve from R1-32B (71.2%) to R1-671B (77.8%)
  6. Well-engineered: KV-cache preservation, speculative caching, implicit compression — practical deployment considerations

Weaknesses / Questions

  1. Step segmentation dependency: The classifier (94.2% F1 in-domain, ~91% OOD) introduces a failure mode — mis-segmented steps propagate errors through RSUS
  2. Verbalized confidence quality: The proxy for completion-only models achieves 0.71 AUROC vs 0.83 for true verbalized confidence — a significant gap
  3. Limited to multi-hop QA: Evaluation only on MuSiQue, HotpotQA, 2Wiki. Generalization to other reasoning tasks (math, coding, planning) unknown.
  4. No comparison to concurrent "Dynamic Search-R1": The paper mentions it as orthogonal but doesn't experiment with combining adaptive timing + adaptive depth.
  5. Fixed k=3 for consistency sampling: No ablation on the number of alternative continuations sampled.
  6. QueryGen is simple: Single-layer transformer decoder — more sophisticated query formulation might yield further gains.

Implications for Our Work

  • Step-level uncertainty detection is directly applicable to agent reasoning: knowing when an agent needs to retrieve vs. reason from memory is crucial
  • The RSUS triple-signal approach (self-report + entity coverage + consistency) is a practical uncertainty quantification method for black-box models
  • Speculative caching (37% hit rate) is a general technique applicable to any agent system with predictable query patterns
  • The granularity insight (reasoning-step > sentence > token) suggests agent memory retrieval should operate at decision-boundary granularity
  • For retrieval systems, this work demonstrates that fewer, better-timed retrievals outperform frequent fixed-interval ones
  • retrieval — Retrieval-augmented generation
  • reasoning — Chain-of-thought and multi-step reasoning
  • agent — AI agent architectures
  • inference — Inference optimization techniques