Revisable Evidence Spans for Multi-Hop Retrieval

Multi-hop questions ask a system to chain facts. For example, "In what city was the director of Inception born?" requires first finding the director and then finding his birthplace. Current retrieval-augmented systems, meaning language models that look up passages before answering, handle this by searching repeatedly. They face a dilemma about how much text to keep. If they keep every retrieved passage whole, the context fills with distractors. If they trim passages to the sentences that look relevant to the current question, they discard facts whose relevance only appears one hop later. The sentence naming Christopher Nolan's hometown looks like noise until Nolan has been identified. Once trimmed, it is gone.

We propose treating every kept piece of evidence as a revisable span: a start and end position inside its source document. At each reasoning hop, a small refinement model re-reads each span with some surrounding text and moves its boundaries outward or inward, conditioned on what the reasoner now knows. Earlier trimming decisions become provisional. The design borrows directly from temporal action detection in video. That field learned to localize events by cascading boundary regressions, where each refined window yields cleaner features for the next adjustment. Here the "event" is the evidence a hop needs, and the "video" is the retrieved document.

Core Claim

The claim has two parts, and each can be tested on its own. The diagnostic part says that a measurable share of multi-hop failures in pruned-context systems come from bridge pruning: a gold supporting sentence was in the retrieved pool at some hop but missing from the reader's context when it produced the answer. The mechanistic part says that re-estimating contiguous span boundaries at every hop recovers these bridges at roughly the token cost of aggressive pruning. The alternative mechanism is re-classifying isolated sentences. The span design should win because contiguity also preserves antecedents, such as the sentence that says who "he" is. A bounded cascade of two to three refinement passes per hop should capture most of the gain, mirroring the plateau seen in video.

Where This Sits

Three lines of prior work motivate the proposal by exposing the gap.

Interleaved retrieval, as in IRCoT, alternates chain-of-thought steps with searches and appends whole paragraphs. It raises recall by 11–21 points, but its buffer grows without bound. CoRAG fine-tunes an 8B model on rejection-sampled retrieval chains and lifts MuSiQue retrieval recall from 29.0% to 47.1%, yet it still ingests retrieved passages in full. FrugalRAG uses reinforcement learning (Group Relative Policy Optimization) to decide when to stop searching. It controls context only at the granularity of whole search turns, under a hard six-hop cap. None of these systems changes what is kept inside a retrieved document.

Compression methods constrain the design. EXIT classifies sentences as keep or drop in one pass with a 2B classifier and reports +3.7 EM on average with a Llama-3.1-70B reader, along with a 2.4x latency cut. This shows that pruning helps even on multi-hop data, so any proposal must match its efficiency. EXIT never revisits a dropped sentence. ECoRAG appends evidence until a sufficiency test passes, which also only moves in one direction. BAR-RAG uses "boundary" to mean the generator's competence boundary, a training-time notion. Its selector is discarded at inference, so it is complementary rather than competing.

Boundary refinement supplies the method. Cascaded Boundary Regression re-crops the video at each refined boundary and regresses a new offset in absolute unit coordinates. This raised THUMOS-14 [email protected] from 19.0% to 31.0%, with gains plateauing after about three steps. Xie et al. push this to 44.2% and show that order-preserving pooling over the proposal matters. We reuse the unit-coordinate regression, the re-crop-and-re-encode loop, and the expected depth ceiling.

The closest agentic work is video-based. VideoTemp-o3 refines one temporal window before answering. SCOUT shows that zoom-in agents fail from "irreversible early commitment" and adds backtracking, but only over a discrete toolset without boundary regression. Our proposal is the text multi-hop analogue that combines both properties: boundaries that are refined continuously, and refinement that is not restricted to zooming in. It applies to all spans in the buffer, not only the newest one.

Research Program

The system has three parts.

The first is a standard interleaved reasoner, Llama-3.1-8B-Instruct in an IRCoT-style loop. The second is a span buffer. Each entry stores a document identifier and sentence-level start and end indices. Pruned material is never deleted from the buffer's source pointer, so any span can later be regrown.

The third is a boundary refiner, a Gemma-2B-scale model matched to EXIT's classifier size. At hop $t$, for each span it reads the current span plus a margin of $m$ sentences on each side, together with the reasoning state $s_t$ (question, sub-queries, and sub-answers so far). It outputs an update:

$$b^{(k+1)} = b^{(k)} + \Delta_\theta\big(\text{doc}[b^{(k)} \pm m],\, s_t\big),$$

where $b^{(k)}$ is the (start, end) pair in sentence units after refinement pass $k$, and $\Delta_\theta$ is an integer offset pair or a "drop" action. As in cascaded boundary regression, each pass re-encodes the text under the new boundaries. The number of passes per hop is capped at $K$, with $K \in \{1,2,3,4\}$ studied explicitly.

Supervision comes from existing annotations. HotpotQA and 2WikiMultiHopQA provide supporting-fact sentences, and MuSiQue provides per-sub-question supporting paragraphs. The regression target for a span is the minimal contiguous sentence range covering the gold supporting sentences in that document, extended to include any sentence containing the antecedent of a pronoun in them. We resolve antecedents with an off-the-shelf coreference pass over training data only.

Targets are presented at every hop, given the state at that hop. The refiner therefore learns to keep "bridge-shaped" content, such as named entities of the type the pending sub-question asks for, even when that content does not match the surface of the question. MuSiQue's decompositions let us also train hop-specific targets, and we compare the two target schemes.

On the analysis side, we instrument every system with the bridge pruning rate. Its denominator is questions where a gold supporting sentence entered the candidate pool. Its numerator is the subset where that sentence was absent from the final reader context. We break the rate down by hop index.

Baselines

The naive reference is single-step retrieval of the top five passages, fed whole to the same 8B reader. CoRAG reports the retrieval-recall level of this setting at 29.0% on MuSiQue. It anchors how much any iterative method helps.

The full-accumulation reference is IRCoT with the same reader and retriever, appending full paragraphs each hop. It has zero bridge pruning by construction, so it shows the accuracy we would get if context bloat cost nothing, alongside its token count and latency.

The strongest existing multi-hop systems are FrugalRAG and CoRAG. FrugalRAG 7B reports 68.47% model-based evaluation (MBE) accuracy on HotpotQA with 2.05 searches on average, against ReAct's 64.20% with 2.76 searches. CoRAG-8B reports 47.1% MuSiQue retrieval recall and EM more than 10 points above IterDRAG and Search-o1-32B. No absolute CoRAG EM value was found on Lacuna. We run FrugalRAG's released policy with its retriever (ColBERT or E5) and our span buffer bolted on, so the question becomes whether revisable spans add accuracy per token on top of learned stopping.

The strongest compression baseline is EXIT applied to each hop's new passages. EXIT's paper reports only the +3.7 EM average and the 8.4s→3.5s latency, and no HotpotQA-specific number was found on Lacuna. We therefore rerun it ourselves.

The key ablation is re-scored EXIT. At every hop, EXIT's classifier is re-run over all sentences of all documents seen so far, conditioned on the cumulative reasoning state. This is the simplest version of "pruning is provisional." It isolates whether our gain comes from revisability alone or from contiguous boundary regression with antecedent preservation. We also run our method with revisiting disabled (only the newest hop's spans are refined), with $K=1$ (no cascade), and a heuristic that keeps each EXIT-selected sentence plus one neighbor on each side.

Fairness rules apply to every comparison. All systems share the corpus, retriever, reader, and hop cap, and they receive equal tuning budgets for thresholds on the same dev split. Comparisons are reported at matched reader-context token budgets, sweeping each method's threshold to trace an accuracy–tokens curve.

Beating the field means two things. At equal tokens, our curve must dominate re-scored EXIT on MuSiQue and 2WikiMultiHopQA by a margin exceeding seed variance. Our method must also match or exceed full-accumulation IRCoT accuracy at no more than half its tokens.

Why not just…?

Why not just give a long-context reader everything? The evidence says long contexts are not free. EXIT improves EM over uncompressed context even with a 70B reader. BAR-RAG reports degradation as distractor count grows. Video-MME documents a 14.3% drop for Gemini 1.5 Pro from short to long inputs. Our full-accumulation IRCoT baseline measures this directly in our setting. If it matches our accuracy at similar latency, the proposal's value reduces to cost savings, and we will say so.

Why not just re-run a sentence classifier every hop with the updated query? This is the real threat, and we adopt it as the central ablation rather than argue it away. Re-scoring already makes pruning revisable. Our bet is that the remaining errors are antecedent losses, which isolated sentence decisions cannot see and contiguous span regression can. The pronoun-antecedent slice of the bridge pruning analysis settles it. If re-scored EXIT closes the gap, we will report that revisability, not boundary regression, is the operative ingredient.

Why not just retrieve larger or overlapping chunks? Larger chunks trade bridge recall for distractor load at a fixed ratio. That is the static-window regime whose weakness on unaligned boundaries is quantified by cascaded boundary regression's 19.0%→31.0% jump over sliding windows. The neighbor-expansion heuristic and chunk-size sweeps on the accuracy–tokens curve test this directly.

Why not just learn when to stop searching? FrugalRAG shows learned stopping cuts searches substantially. However, it cannot change what is inside a retrieved passage, and the two ideas compose. Running FrugalRAG with and without the span buffer tests whether our method adds anything once search count is already optimized.

Decisive Checks

The first check is existential. If the bridge pruning rate under single-pass EXIT on MuSiQue and 2WikiMultiHopQA is small, meaning the gold-context oracle recovers little accuracy, the premise fails and we stop.

The second check is the matched-token comparison against re-scored EXIT, sliced by questions whose supporting sentences contain unresolved pronouns. This is where the span mechanism must show its advantage.

The third check is the cascade-depth curve. We predict gains saturate by $K=2$–3 and degrade beyond, echoing the video results. Continued improvement instead would suggest the CoRAG-style log-linear scaling with test-time compute also applies at the span level, which would be interesting in itself.

The fourth check is robustness. Following BAR-RAG's protocol, we inject increasing numbers of distractor passages and ask whether revisable spans degrade more slowly than whole-passage accumulation.

Evaluation covers HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle. We report EM/F1, MBE (for comparability with FrugalRAG), supporting-fact recall in the final context, reader tokens, latency, and bridge pruning rate.