acceptodds
Under review as a conference paper at ICLR 2027

From Delivery to Stateful Exploration: Rethinking the Index for Agentic Search

33 views

Abstract

Recent advances in agentic search have given large language model (LLM) agents finer control over corpus exploration. However, search interfaces often return matching passages even when feedback about the candidate set would suffice for the next decision, coupling candidate refinement with source-text exposure. We propose IndexAct, an interface for Index-Native Corpus Interaction that separates candidate-set refinement from text inspection. Agents construct and manipulate persistent candidate sets through lexical conditions and set operations over an inverted index, receiving reusable state references and statistics such as candidate counts rather than matching passages. This feedback guides further refinement, while separately requested passages provide new clues or evidence that can inform subsequent operations on retained candidate sets. Experiments on five benchmarks spanning agentic search and multi-hop question answering show that IndexAct outperforms the evaluated baselines on each benchmark. On BrowseComp-Plus, it also achieves higher evidence coverage with a smaller average live context than terminal-based corpus interfaces, and maintains answer accuracy as the corpus expands. Further analyses suggest that informative refinement feedback and state reuse support continued evidence discovery, while shorter contexts or fewer search steps alone do not ensure better performance. Code: https://anonymous.4open.science/status/IndexAct

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, agentic search, information retrieval, search interfaces, index-native corpus interaction, context management

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bxU4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces INDEXACT, an interface for agentic search that separates candidate-set exploration from source-text inspection. Motivated by an analysis of context composition and the timing of source-text exposure, the method enables agents to construct, retain, and combine candidate sets through index-based operations. These operations return reusable state references and statistical feedback, while source text is accessed through explicit reading requests.

Critical strengths
  1. The progression from empirical observations to problem formulation and method design is clear and well motivated.

The paper begins by examining what enters an agent’s context and when source text is encountered during evidence discovery. It then asks a concrete question: does the agent’s next search decision require reading source text, or would feedback about the candidate set suffice? This question naturally motivates the proposed combination of persistent candidate states, statistical feedback, and explicit reading. The resulting argument is coherent, with the method’s components closely connected to the identified problem. The authors also appropriately acknowledge that text from unannotated documents may still provide useful clues.

  1. The main benchmark results are promising, with reasonably broad comparisons and component-level analyses.

On the full 830-question BrowseComp-Plus benchmark, INDEXACT improves answer accuracy and evidence coverage while reducing average live context relative to the terminal-based alternatives. The evaluation includes conventional retrieval agents, direct corpus interaction methods, and selective-reading interfaces. The additional multi-hop QA experiments broaden the evaluation, while the ablations provide initial evidence that exact counts, cross-call state reuse, and set composition contribute to performance.

Critical weaknesses
  1. Although the paper includes several relevant methods, important details remain unclear, including the evaluated RARG variant, PI-SERINI’s retrieval depth and BM25 configuration, and the adaptation of SIEVE’s structured-document interface to the evaluation corpora. These choices could materially affect performance. Without clearer documentation, it is difficult to determine how much of the observed improvement can be attributed specifically to the proposed interaction mechanism.

  2. Some improvements on multi-hop QA are very small. The gains of 0.2 percentage points on 2Wiki and 0.8 percentage points on Bamboogle each correspond to only one additional correctly answered question. The main ablations use a 100-question subset, and the paper does not report repeated-run variability, confidence intervals, or paired significance tests. Consequently, the stability of the smaller improvements and some ablation differences is difficult to assess. Furthermore, all experiments use a single backbone model. While the BrowseComp-Plus result is encouraging, additional statistical analysis and evaluation with another backbone would strengthen the broader claim that the proposed interface consistently improves agentic search.

Questions for authors

see above

Reviewer_nsRYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

INDEXACT separates candidate-set refinement from source-text reading. Agents manipulate persistent search states using Boolean operations, counts, and positional constraints, then selectively inspect passages. Experiments report improved accuracy across five benchmarks and reduced live context compared with terminal-based interfaces.

Critical strengths
  1. The interface offers a clear, practical contribution: reusable candidate states let agents explore alternatives without repeatedly loading source text.
  2. Results and ablations support the design. On BrowseComp-Plus, accuracy improves from RARG’s 69.4% to 73.5%, while average live context roughly halves; removing state reuse or set composition reduces performance.
Critical weaknesses
  1. Some benchmark wins amount to one additional correct answer, notably 2Wiki and Bamboogle. No confidence intervals or repeated runs are reported, and all experiments use one backbone.
  2. Context comparisons are difficult to reproduce because truncation/compaction policies and run-wide reading limits are not clearly specified across interfaces.
  3. Appendix E discusses backend latency, timeouts, and state storage without numerical results. Stable context and API cost therefore do not establish overall computational efficiency.
Questions for authors
  1. Are the small QA improvements supported by paired confidence intervals or repeated runs?
  2. Which context-management policies and reading limits were used for each interface?
  3. What are backend latency and peak state-storage costs as the corpus grows?
Reviewer_ct6UOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes IndexAct, a framework that provides a retrieval agent (ReAct style) with retrieval primitives instead of just a retrieval model. It allows the agent to collect statistics about terms, intersections, proximities, etc, before actually retrieving the document tokens into the tokens to verify and refine the next query turn. The idea is interesting and promissing, as well as the results.

Section 3 provides a motivation that is not fully convincing or needed for that matter - the idea is simple and straighforward. Section 4 is clear and nice to read. Experiments are convincing, but not in-depth enough. There are many abblations that would strengthen the idea, e.g. other patterns beside ReAct, other base models.

My main criticism concern the scope and the target audience.

Critical strengths
  • The idea of decomposing retrieval primitives and providing an agent with them is an interesting and novel idea.
  • It is a training free approach as long as the base LLM has strong planning and reasoning capabilities.
  • Evaluation is extensive and convincing. However, it doesn't look quite competitive against other specialized baselines for multi-hop QA.
Critical weaknesses
  • Section 4 provides a series of basic primitives. However, there seems to be many possibilities and the current proposal only scratches the surface.
  • Evaluation is limited to GPT-5.4nano. It would be interesting to see how the idea generalizes to other open weight models.
  • The scope of the contribution seems to be better suited to an Information Retrieval audience.
Questions for authors
  • Analysis 1: How many documents were selected? How many tokens are in each case (BM25 and DCI)? Why not the top 10 docs with BM25? Which dataset is this? How many queries? How many relevant documents per query?
  • Analysis 2: Isn't the second and following rounds of retrieval impacted by the trajectory queries? How were they decided?
  • Would it be possible to collect statistics such as frequency of terms proximal to the query terms in the retrieved set? This could be relevant to the next query turn. There seems to be many possibilities and the current proposal only scratches the surface.
  • The maximum of 300 query turns seem to be too many. How many query turns per QA are made on average? How many index inspections are done per QA? How many document retrieval operations are done per QA on average? How does it compare to the others?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.