acceptodds
Under review as a conference paper at ICLR 2027

CLOSED-LOOP BRIDGE RETRIEVAL FOR MULTI-HOP QUESTION ANSWERING

Abstract

Multi-hop retrieval-augmented generation has to connect supporting evidencescattered across documents, and the right next retrieval target usually becomesclear only after earlier evidence accumulates. Iterative RAG systems adapt re-trieval through query rewriting, subquestion generation, or sufficiency assessment,but this adaptation stays open-loop at the bridge that routes the next hop: a candi-date intermediate entity steers the following retrieval step, and whatever evidencecomes back is folded into the context without asking again whether the bridge–gapbinding behind it is actually supported, so a wrong route is carried forward insteadof revoked. What remains open is when retrieval should commit to an interme-diate bridge and how it should recover once later evidence contradicts that com-mitment. We propose CLBR-RAG (Closed-Loop Bridge Retrieval), which closesthe loop by turning each bridge commitment into a provisional control state thatcan be revoked. A commitment binds an evidence-supported intermediate entityto an unresolved information gap. Complementary lexical and semantic querieskeep coverage over candidate bridges before any commitment is made; the Ev-idence Gate admits a source-grounded binding only when it matches an activegap and agrees with the accumulated context, and a deterministic composer theninstantiates its targeted next-hop query. The evidence returned by that branchdecides whether the commitment survives: unsupported branches are retracted,supported ones keep their evidence, and BROAD exploration resumes when notargeted branch is accepted. Validated cross-document bridge relations are car-ried into a compact, provenance-preserving context for final reasoning. Acrossthree multi-hop benchmarks and TriviaQA as a broader control, CLBR-RAG isconsistently effective. Code will be released upon publication.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Retrieval-Augmented Generation, Multi-Hop Question Answering, Evidence-Grounded Retrieval, Retrieval Control, Adaptive Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_whXNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem in multi-hop RAG where incorrectly selected intermediate entities lead subsequent retrieval down the wrong path. The proposed CLBR-RAG represents the combination of an intermediate entity and an unresolved information need as a tentative retrieval control state, and approves or retracts it through a pre-retrieval Evidence Gate and a post-retrieval support check. In the retrieval stage, it combines BM25 and E5 results with RRF, falls back to BROAD search when all branches of goal-directed retrieval fail, and provides compressed evidence and verified linking relations to the final reasoning stage. It reports the highest EM/F1 among comparison methods on the full development sets of four benchmarks, with especially notable gains of 5.0/6.5%p over STRIDE on MuSiQue. Ablation studies and retrieval-path diagnostics support the practical value of the design, but important supplementary evidence is needed regarding the specificity of the core verification rules, the scope of retractable states, and experiments that demonstrate the effectiveness of the closed-loop control itself.

Critical strengths

S1. Consistent performance improvement on various datasets. The paper shows higher EM/F1 scores than other baselines on four datasets. In particular, it is notable that the proposed method shows significant improvement (+5.0 pp for EM and +6.5 pp for F1) compared to the strongest baseline.

S2. Ablation studies. It shows the sufficient ablation studies in terms of retrieval, bridge control, and evidence organization. Furthermore, it provides the pass rate for each stage, rejection cause, etc. It is positive that the authors tried their best to analyze on which stage binding and branch are accepted or rejected.

Critical weaknesses

W1. Lack of implementation specification of core verification function. Evidence gate and post-target validation are the deterministic program check without additional model call. However, there is not sufficiently described implementation in terms of DependencyWitness, NoBindingConflict, and target relation or witness judgement of slot in Support.

W2. Narrow revocation scope in implementation side compared to author's claim. Algorithm 1 discards provisional branch failed validation after TARGETED retrieval. In contrast, it does not provide the procedure to remove the existed binding when contradictory evidence was found under the case that binding is accumulated (Ht+1=Ht∪Pt+1+H_{t+1} = H_t \cup P_{t+1}^{+} ) after it passes validation and goes to the ledger. Therefore, the proposed method shows discard of validation-failed provisional branch, not general recovery that revokes the wrong bridge commitment that is already accepted. The authors should constraint the claim of this paper.

W3. The correctness of downstream evidence does not guarantee the validity of upstream bridge. The post-target validation in Eq. 12 confirms whether the returned evidence supports the bridge entity and target relation or slot. In terms of director, however, it can retrieve the exact birthplace evidence. With the fact that confirms the birthplace of corresponding entity, the upstream relation is hard to be justified that that entity the film director in the question.

W4. Contribution of recovery. In Table 5, the additional improvement by adding full control after pre-target Gate is +0.3 pp for EM and +0.5 pp for F1. In this comparison, there is no separation between post-target validation and the effect of BROAD recovery, so it is hard to show the contribution of recovery component. It requires comparison with three splitted configuration: (1) Only Pre-target Gate, (2) Pre-target Gate + post-target validation without BROAD recovery, (3) Pre-target Gate + post-target validation + BROAD recovery.

W5. Needs of comparison constraint with retrieval and entire inference cost. The proposed method utilized BM25-E5-RRF hybrid retrieval. Therefore, to show the additional contribution of bridge control, it should be applied to identical retrieval backend in the strong baselines.

Questions for authors

See the weaknesses, and I will write this section as minor comments.

1. Missing references. The authors missed the references such as datasets (TriviaQA, MuSiQue) and metrics (BM25, E5, RRF).

2. Location of figures and tables. Typically, figures and tables should be located at the top of the paper.

Reviewer_DfGSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLBR-RAG to address retrieval path errors caused by incorrect intermediate bridge entities in multi-hop RAG. The framework consists of three components: complementary evidence retrieval to preserve candidate coverage, closed-loop bridge control to guide targeted retrieval and validate evidence, and evidence organization to compress and integrate multi-round evidence. Experimental results demonstrate the effectiveness of the proposed approach.

Critical strengths

The paper identifies an important failure mode in multi-hop RAG, where premature commitment to incorrect bridge entities propagates errors across retrieval steps. Modeling bridge commitments as revocable states is a reasonable design, and the combination of pre-retrieval verification, post-retrieval validation, and failure recovery provides a coherent closed-loop framework.

Critical weaknesses

W1. Limited methodological novelty. The main components of CLBR-RAG, including hybrid retrieval, gap assessment, evidence verification, retrieval fallback, and context compression, are closely related to existing RAG techniques. In particular, the Base Judge and basic evidence compression are directly adapted from S2G-RAG. Although modeling bridge commitments as revocable states offers some novelty, the paper does not sufficiently establish its fundamental distinction from existing iterative retrieval and self-correction methods.

W2. Insufficient evaluation of bridge verification reliability. The Evidence Gate relies primarily on entity string matching, gap compatibility, and history consistency, which may be insufficient for verifying complex semantic relations. The paper does not directly report the precision or recall of bridge candidates proposed by the Anchor Judge, nor does it analyze the false rejection of valid bridges. Consequently, the reliability of the core bridge verification mechanism remains unclear.

W3. Although some retrieval budgets and generation models are standardized, CLBR-RAG introduces multiple judges, hybrid retrieval, and additional evidence processing. The reported improvements may partly result from these additional components rather than the proposed closed-loop bridge control. Stronger baselines under identical BM25–E5 retrieval settings and matched LLM call budgets are needed.

W4. The robustness of CLBR-RAG to entity ambiguity, noisy or conflicting evidence, and incomplete knowledge bases remains insufficiently explored.

Questions for authors

See Weakness.

Reviewer_WKYWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CLBR-RAG treats an intermediate entity–gap binding as a provisional retrieval decision rather than a fixed route. BM25/E5 broad exploration supplies candidates; an Anchor Judge proposes bindings, deterministic checks admit them, and targeted evidence either validates or retracts each branch. Broad exploration resumes when no branch is accepted. The reasoner receives compact evidence with provenance and validated bridge relations. The system reports improvements across four QA benchmarks, with particular emphasis on multi-hop retrieval.

Critical strengths

The branch-level commitment and recovery design is a coherent operational idea, and preserving source pointers makes the intended evidence flow easy to follow. The full-development-set improvements over the strongest baselines are encouraging: 2.2/3.6 EM/F1 points on HotpotQA, 4.1/2.4 on 2Wiki, and 5.0/6.5 on MuSiQue. Comparisons share the answer-generation backbone and retrieval-round/document budgets. The shared-prefix replay is a useful controlled diagnostic, and the appendix adds latency, alternative-backbone, and model-capacity results.

Critical weaknesses

My main concern is that the central deterministic checks remain underspecified. Equations 7–8 name DependencyWitness and NoBindingConflict, while the post-target check requires a relation/slot witness and branch coverage. Appendix A.3 largely restates these requirements. Although Appendix C supplies model prompts, it does not explain how the program recognizes open-vocabulary relations, normalizes aliases, or detects contradictions. I therefore cannot tell how semantic support is distinguished from entity/slot co-occurrence, which limits reproducibility and the meaning of a “validated” binding.

The evidence for the specific post-retrieval validation and recovery mechanism is also limited. Table 5 moves from 48.9/58.9 EM/F1 with the pre-target gate to 49.2/59.4 with full control, without uncertainty estimates or separate validation-versus-recovery ablations. Stage-wise retention measures the controller’s acceptance decisions, not independently established bridge correctness. The acknowledged absence of proposal precision/recall reinforces this gap. Strong end-to-end scores support the overall system, but do not yet establish that its verifier reliably detects and repairs incorrect commitments. I lean toward weak rejection until these two points are clarified.

Questions for authors

Could you provide executable definitions of DependencyWitness, NoBindingConflict, and TargetedSupportCheck, including worked examples with aliases, paraphrased relations, and negation?

Could you isolate post-target support validation from BROAD fallback, report paired confidence intervals, and audit accepted/rejected bridges against gold annotations or human judgments to quantify incorrect-bridge acceptance, rejection, and subsequent error recovery?

1 comment
AuthorsOpenReview ↗

Executable Verification Rules

Response to WKYW — Executable Verification Rules

Thank you for requesting executable definitions. We clarify Eqs. (7)–(8), Eq. (12), and Appendix A.3.

1. Relation witnesses and semantic matching

Each expected dependency is a directed (head, relation, tail) triple. Fixed active/passive/role frames match the predicate cue and both argument spans, not word co-occurrence. For developer(StreamKit, Orion Devices), the registered patterns include:

ACTIVE:  <tail> developed <head>
PASSIVE: <head> was developed by <tail>
NEGATED: <head> was not developed by <tail>

The first two yield positive witnesses; the third yields a negative one. Source-backed aliases (e.g., Orion Devices (OD)) require an explicit supporting span. Each witness records its source ID, head/cue/tail spans, rule ID, and polarity. Conditional, uncertain, or non-asserted expressions cannot establish positive support.

Evidence aggregation is: positive only → SUPPORTED; negative only → REFUTED; both → CONFLICTED; neither → UNKNOWN. Unregistered paraphrases are not presumed equivalent.

2. Executable decision logic

The following pseudocode specifies the core acceptance and rejection conditions:

DependencyWitness(e, gap, src, q):
  if invalid_source(src) or not source_authorized(e,src):
    return UNKNOWN
  D = compile_dependencies(q,gap,e)
  if not gap_compatible(D,gap): return UNKNOWN
  V[d] = verdict(directed_matches(d,src)) for each d in D
  P = dedup({d : V[d] == SUPPORTED})
  if |P| > 1: return UNKNOWN
  if |P| == 1: return SUPPORTED + source/span receipt
  if any V[d] in {REFUTED,CONFLICTED}: return negative verdict
  if restricted_direct_identity(q,gap,e): return SUPPORTED
  return UNKNOWN

NoBindingConflict(e,gap,history,witness):
  if not replay_upstream(witness,e): return UNKNOWN
  for each active h with matching head and canonical relation:
    if h.receipt exists but fails replay:
      mark unresolved; continue
    potential = (same_tail(h,witness) and negative(h)) or
                (different_positive_tail(h,witness) and
                 gap.binding_cardinality == 'one')
    if potential and overlapping_time(h,gap): mark conflict
    if potential and time_unknown(h,gap): mark unresolved
  if conflict: return CONFLICTED
  if unresolved: return UNKNOWN
  return SUPPORTED

TargetedSupportCheck(b,evidence,active_gaps,upstream,history):
  require nonempty branch_id, gap_id and active_gaps
  require replay_upstream(upstream,b.entity)
  require matching parent gap, active gap and NBC == SUPPORTED
  # Failed prerequisites cannot pass; propagate conflicts.
  required = refine_required_slots(b, original_question)
  for s in evidence with valid source and matching branch/gap IDs:
    if s contradicts upstream dependency: return CONFLICTED
    collect slot_witness(b.entity,slot,s) for each required slot
  if any slot verdict == CONFLICTED: return CONFLICTED
  if all required slots == SUPPORTED: return SUPPORTED
  return UNKNOWN

replay_upstream checks rule version, receipt integrity, recorded source and relation witnesses, and re-executes the upstream test; a hash alone is not semantic proof. Direct identity is restricted to explicitly named subjects in direct-attribute questions, not unresolved intermediate bridges.

3. Worked positive and negative examples

For “Where are the headquarters of the developer of StreamKit?”, source d:s0:

“Orion Devices (OD) is a company. StreamKit was developed by OD.”

This yields a positive developer(StreamKit, Orion Devices) witness (developer:passive, head span [33,42], cue [47,59], tail [60,62]). Thus DW=SUPPORTED; with empty history, NBC=SUPPORTED.

A same-branch source “Orion Devices is headquartered in Berlin” supports the required headquarters slot, giving TSC=SUPPORTED.

Conversely, “StreamKit was not developed by Orion Devices” yields REFUTED; “Harbor Labs developed StreamKit; Orion Devices was mentioned separately” yields UNKNOWN for the proposed dependency despite lexical co-occurrence.

An overlapping negative historical binding yields CONFLICTED; distinct positive developers conflict only under an explicit single-value constraint.

Crucially, a correct birthplace fact for a candidate director cannot validate the film–director dependency without an independent upstream witness. Evidence from another branch or gap cannot satisfy the current target.

4. Validation scope

Only supported branches and their evidence enter the ledger. Failed provisional branches are discarded; all-branch failure triggers BROAD recovery.

Thus, a validated bridge satisfies explicit, source-traceable upstream and downstream conditions rather than guaranteeing unrestricted factual correctness. Semantic coverage is bounded by registered linguistic patterns and typed extractors; unrecognized cases conservatively return UNKNOWN.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.