PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Information Control for Search-Augmented LLM Reasoning

Siheng Xiong, Oguzhan Gungordu, James C. Kerce, Faramarz Fekri

arXiv:2602.01672v2cs.CL

TL;DR

Search-augmented reasoning lacks explicit adaptive control over information acquisition, while outcome-based RL offers limited guidance for intermediate retrieval decisions. DeepControl uses utility-driven training signals and annealed control to regulate retrieval extent and resolution. Across seven datasets, it consistently outperforms strong baselines, improving average EM over Search-R1-instruct by 9.4 points on Qwen2.5-7B and 8.6 points on Qwen2.5-3B.

  • Problem

    Outcome-based RL provides limited guidance for intermediate retrieval decisions, and explicit adaptive control over acquisition extent and resolution is missing.

    Source: §1 · p. 2

  • Method

    DeepControl uses utility-driven training signals and annealed control to regulate retrieval extent and resolution, helping policies internalize acquisition behavior during online RL.

    Source: §1 · p. 2

  • Results

    Across seven datasets, DeepControl consistently outperforms strong baselines, improving average EM over Search-R1-instruct by 9.4 points on Qwen2.5-7B and 8.6 points on Qwen2.5-3B.

    Source: §4.2 · p. 6

  • Takeaways & Limitations

    The results indicate that effective search-augmented reasoning requires controlling when and how retrieved information is used.

    Source: §4.2 · p. 6

    The evaluation focuses on text-based question-answering benchmarks and does not study dynamic corpora, multi-tool or multi-agent settings, or multimodal reasoning.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Search-augmented reasoning agents interleave multi-step reasoning with external retrieval, but uncontrolled retrieval can introduce redundant evidence, saturate the context, and destabilize reinforcement learning (RL). Existing outcome-based RL methods provide only sparse terminal rewards, offering limited guidance for intermediate information-acquisition decisions. We propose DeepControl, an adaptive information-control framework based on information utility, a state-dependent estimate of the marginal value of retrieved evidence. The framework regulates information acquisition along two axes: extent, i.e., whether retrieval should continue, and resolution, i.e., how much retrieved detail should be exposed. It implements these controls through retrieval-continuation guidance, hierarchical granularity control, and an annealed control-forcing scheme. This enables the policy to internalize effective acquisition behavior during training and operate without external control at test time. Across seven benchmarks, DeepControl consistently outperforms strong RL and retrieval baselines without explicit information control; compared with Search-R1, it improves average performance by +9.4 and +8.6 points on Qwen2.5-7B and Qwen2.5-3B, respectively. Additional analyses show improved search effectiveness, training stability, and evidence utilization.

1 Introduction

DEEPCONTROL treats information acquisition as a unified control problem, regulating both whether to retrieve more evidence and how much detail to expose. It uses information utility and annealed training-time guidance to help search-augmented reasoning agents learn these behaviors.

1 Introduction

  • Outcome-only RL gives limited guidance for intermediate retrieval decisions, leading agents to over-retrieve or stop prematurely; more retrieval can also degrade decision quality through saturation and noisy evidence.These problems are reported for long-horizon settings and weak evidence or poorly specified queries (Xiong et al., 2026, 2025b).Source: §1 · pp. 1–2
  • DEEPCONTROL frames information acquisition along two axes: extent, or whether to continue retrieval, and resolution, or how much retrieved detail to expose.This unifies control of retrieval continuation and evidence granularity rather than treating stopping, retrieval quality, and context construction separately.Source: §1 · p. 2
  • Information utility estimates the state-dependent marginal value of retrieved evidence by combining novelty and effectiveness.It serves as a practical signal for information-acquisition control during training.Source: §1 · p. 2
  • DEEPCONTROL uses utility-driven guidance for retrieval continuation and hierarchical granularity control, with annealed intervention to help the policy internalize effective acquisition behavior.Continuation control addresses premature stopping and over-retrieval; granularity control selectively expands high-utility content.Source: §1 · p. 2
  • Experiments across diverse search-augmented reasoning benchmarks show consistent improvements in reasoning accuracy, training stability, and evidence utilization.Source: §1 · p. 2

2 Preliminaries

The section formalizes a search-augmented reasoning agent that alternates reasoning and retrieval, then outlines online RL and the limitations of outcome-only supervision for controlling information acquisition.

2.1 Problem Formulation

  • The agent samples structured actions containing reasoning tokens, an action type, and parameters, forming trajectories that end with a final answer or at a step limit.Its state includes accumulated context, retrieved evidence, and intermediate reasoning.Source: §2.1 · p. 2

2.2 Online RL with Search-Augmented Reasoning Agents

  • Online RL alternates policy-generated rollouts and policy updates to maximize task success while regularizing deviation from a reference policy.Source: §2.2 · p. 2
  • PPO uses a clipped surrogate objective with GAE advantages and a value function, whereas GRPO estimates advantages within response groups without learning a value function.PPO (Schulman et al., 2017) uses GAE (Schulman et al., 2015); GRPO is regularized toward a reference policy (Shao et al., 2024).Source: §2.2 · pp. 2–3
  • Outcome-based RL can leave intermediate retrieval decisions unguided, leading agents to over-retrieve, accumulate unnecessary context, and rely on sparse final-answer correctness rewards.Policy-gradient updates apply only to language-model-generated tokens, while Exact Match evaluates final-answer correctness.Source: §2.2 · p. 3
  • Raw retrieved content can saturate context; small top-k retrieval risks missing evidence, while longer contexts increase training and inference costs.Source: §2.2 · p. 3
  • Sparse outcome-based supervision makes optimization sensitive to errors over long reasoning trajectories, especially when weaker models explore inaccurately.Source: §2.2 · p. 3

3 Adaptive Information Control

DeepControl uses state-dependent information utility to address sparse terminal supervision and guide retrieval extent—whether to continue—and resolution—how much retrieved detail to expose. Its controls support selective evidence acquisition while training the policy to act autonomously at test time.

3.1 Information Utility

  • Information utility is a dense, state-dependent training-time signal that evaluates each search step relative to the current reasoning state, balancing novelty and effectiveness.Each search step consists of a retrieval action followed by optional expansion actions.Source: §3.1 · pp. 3–4, §3 · p. 3

3.2 Granularity Control via Hierarchical Selective Expansion

  • Granularity control injects coarse retrieved roots first, then selectively expands high-utility evidence into finer units to prioritize useful detail while limiting context growth.During training, the method selects top-kexpand leaves by utility and traces their ancestors to form expansion targets.Source: §3.2 · p. 4

3.3 Search Continuation Control

  • Continuation control stops after m_stop consecutive low-utility search steps or forces one extra step when utility stays high but gold-answer confidence is insufficient.Stopping injects a “Stop searching” signal; continuation is one-shot and depends on a gold-answer score threshold.Source: §3.3 · pp. 4–5

3.4 Reinforcement Learning with Information Control

  • An annealed control-forcing curriculum shifts rollouts from frequent external control early to autonomous policy behavior in the final stage.The composite reward retains an F1-based outcome reward with a format floor and penalizes tool-use violations and control non-compliance.Source: §3.4 · p. 5

4 Experiments

DEEPCONTROL improves performance across seven benchmarks and shows gains in training behavior, retrieval control, and robustness analyses.

4.1 Experimental Setup

  • Experiments cover seven general- and multi-hop QA benchmarks and compare retrieval and fine-tuning baselines under matched retrievers, corpora, retrieval budgets, training data, and pretrained models.Training uses Qwen2.5-3B/7B and merged NQ and HotpotQA data; retrieval uses E5 over 2018 Wikipedia, with titles and first sentences as roots and full passages as leaves.Source: §4.1 · pp. 5–6

4.2 Main Results

  • +9.4 average EM points with Qwen2.5-7B and +8.6 with Qwen2.5-3B versus Search-R1-instruct; DEEPCONTROL improves all seven benchmarks across task types.The results indicate that information control benefits retrieval-based reasoning, not merely retrieval access.Source: Table 1 · p. 6, §4.2 · p. 6

4.3 Analysis

  • DEEPCONTROL achieves higher training rewards than vanilla PPO; under annealed control-forcing, PPO remains stable while GRPO degenerates, and autonomous search and expansion rise as control messages fall.These trends suggest that the policy progressively internalizes the desired retrieval behavior.Source: Figure 5 · p. 6, §4.3 · p. 7
  • Search-continuation and granularity controls are complementary, and the tool-usage penalty has the largest effect among reward-related ablations, supporting action-validity regularization.The ablation study also examines control versus reward shaping and hierarchical evidence construction.Source: Table 2 · p. 7, §4.3 · p. 7
  • Later checkpoints achieve higher early-step utility without control, and the default annealing schedule yields the best final no-control performance; overly fast or slow schedules hurt results.The schedule analysis supports gradual internalization of effective search behavior.Source: Figure 6 · p. 7, §4.3 · p. 7
  • Intermediate control values best balance premature stopping and unnecessary continuation, while performance is most sensitive to the tool-usage penalty among reward-design choices.The framework is robust to moderate hyperparameter changes.Source: §4.3 · p. 7
  • Across 200 HotpotQA failures, removing continuation mainly increased insufficient retrieval, while removing granularity mainly increased missing supporting evidence; DEEPCONTROL reduced the main behavioral failure modes.Failures were grouped as insufficient retrieval, retrieval drift, missing supporting evidence, reasoning failure, and format failure.Source: Figure 7 · p. 7, §4.3 · p. 7
  • With BM25, DEEPCONTROL still improves over Search-R1, extending robustness beyond the main E5-retriever setting.The evaluation also discusses model-family selection and the scope of the analysis.Source: §4.3 · pp. 7–8

5 Related Work

Prior retrieval and RL approaches leave information acquisition insufficiently controlled, motivating a unified approach that adapts whether retrieval continues and how much detail is exposed. The section situates this gap among retrieval methods, sparse-reward RL, and prior resource-allocation research.

Large Language Models with Retrieval

Reinforcement Learning for LLM Reasoning and Tool Use

  • Outcome-based RL can induce reasoning from task-level rewards (Shao et al., 2024; Guo et al., 2025), but sparse feedback leaves retrieval decisions weakly guided and tool use brittle.Reported failure modes include premature stopping, over-retrieval, and unreliable tool use.Source: §5 · p. 8

Information Control in Search-Augmented Reasoning

  • Prior work studies novelty, resource allocation, LLM exploration, process supervision, and adaptive retrieval, but addresses acquisition partially rather than jointly controlling extent and resolution.The cited approaches include intrinsic rewards, count-based exploration, curiosity-driven objectives, adaptive retrieval, stopping heuristics, and selective context construction.Source: §5 · pp. 8–9

6 Conclusion

DeepControl uses information utility to regulate search-augmented reasoning, helping models acquire information more effectively without external intervention at test time. Experiments show gains in reasoning accuracy, training stability, and evidence utilization.

6 Conclusion

  • The framework regulates information acquisition through search continuation for extent and hierarchical expansion for resolution.Both controls are based on information utility, a state-dependent estimate of retrieved evidence’s marginal value, as described in the paper context.Source: §6 · p. 9
  • Annealed control-forcing during online reinforcement learning helps models internalize effective information acquisition without external intervention at test time.Source: §6 · p. 9
  • Experiments across multiple tasks and datasets show consistent gains in reasoning accuracy, training stability, and evidence utilization.Source: §6 · p. 9

Limitations

DeepControl is evaluated only for text-based search-augmented question answering, leaving broader settings unstudied. It may still propagate retrieval errors or learn suboptimal behavior, and high-stakes deployment would need further safeguards and evaluation.

Limitations

  • Evaluation is limited to text-based question answering; dynamic corpora, multi-tool and multi-agent settings (Zhang et al., 2025, 2026; Shi et al., 2026), and multimodal reasoning (Li et al., 2026c; Yang et al., 2026a) remain unstudied.Extending adaptive information control to these broader settings is left for future work.Source: Limitations · p. 9
  • The framework may propagate errors from retrieved evidence or learn suboptimal retrieval behavior from imperfect training signals, and high-stakes deployment remains unevaluated.Such use would require additional safeguards and evaluation.Source: Limitations · p. 9

A Information Utility

Information utility estimates the state-dependent marginal value of retrieved evidence for the downstream task. It separates novelty across all retrieved evidence from effectiveness of the evidence actually injected, providing an external signal to regulate information acquisition.

A Information Utility

  • Information utility measures the marginal value of newly acquired information relative to the agent’s current reasoning state and downstream task.The framework uses utility as an external control signal rather than incorporating it directly into the RL reward, keeping utility estimation modular and supporting simpler, more stable RL training.Source: A Information Utility · pp. 13, 15
  • Utility balances novelty over the full retrieved leaf pool with effectiveness of the net injected nodes, detecting redundancy even when retrieved leaves are not expanded.Novelty is measured at concrete leaf-evidence level, while effectiveness can reflect injected internal and leaf nodes.Source: A Information Utility · p. 14
  • Effectiveness measures positive improvement in gold-answer likelihood from newly injected evidence and is zero when confidence fails to improve.It uses a fixed reasoning trace to isolate evidence effects and is available only during training with gold answers; proposed gold-free proxies include entropy, KL-based confidence change, self-consistency, and verifier scores.Source: A Information Utility · pp. 14–15
  • Novelty rapidly decreases after the first retrieval, while later evidence can remain effective in improving answer confidence.This pattern indicates that later evidence may be less novel yet still useful.Source: A Information Utility · p. 15
  • The utility formulation assigns diminishing marginal value to increasingly redundant retrieval after sufficient task-solving evidence has been acquired.Novel and beneficial evidence increases utility, whereas redundant evidence or evidence that does not improve gold-answer likelihood yields little utility gain.Source: A Information Utility · p. 15

B Adaptive Information Control · B.1 Granularity Control via Hierarchical Selective Expansion

Granularity control exposes retrieved evidence coarse-grained first, then lets the agent selectively expand details when needed. During training, utility-based expansion guidance teaches the agent to prioritize valuable evidence while limiting context growth.

B.1 Granularity Control via Hierarchical Selective Expansion

  • Because retrieved information can be lengthy and details vary in usefulness across reasoning stages, the agent starts with coarse results and selectively expands them as needed.Source: §B.1 · p. 15
  • Retrieval and refinement are decoupled: each retrieved source is modeled as a rooted hierarchy whose nodes represent evidence at different resolutions, with edges linking refinements.Source: §B.1 · p. 15
  • After retrieval, only root nodes enter the context; the agent can add expansions until the next retrieval or termination, with the observed evidence set growing incrementally.Source: §B.1 · pp. 15–16
  • An expansion action specifies hierarchy edges from observed nodes to their children, and executing it adds those child nodes to the observed evidence set.Source: §B.1 · p. 16
  • During training, information utility scores retrieved leaf nodes; the top-kexpand leaves and their ancestors define target evidence sets and desired expansion edges.Source: §B.1 · p. 16
  • The model learns to align its expansion actions with the guidance, prioritizing high-utility information while minimizing context growth.Source: §B.1 · p. 16

B.2 Search Continuation Control

Search continuation control uses information utility to monitor the agent’s autonomous search decisions and trigger corrections when it systematically stops too early or searches too long. The learned policy remains the primary decision-maker; utility-based interventions correct specific stopping errors.

B.2 Search Continuation Control

  • Autonomous reasoning can stop prematurely when it undervalues more information or continue searching after useful evidence is exhausted.External intervention is reserved for utility signals indicating systematic misjudgment.Source: §B.2 · p. 16
  • Search stops after information utility stays below δ_stop for m_stop consecutive steps, at which point a control signal explicitly terminates further search.The stopping index is defined by the first qualifying run of low-utility steps.Source: §B.2 · p. 16
  • A one-shot continuation signal adds one search step when recent utility is consistently high but the agent attempts to stop despite insufficient confidence in the gold answer.The continuation criterion uses δ_cont and a gold-answer score threshold; it is applied only during training, when gold answers are available.Source: §B.2 · p. 16

B.3 Reinforcement Learning with Information Control

The method uses control signals to guide early learning, then gradually removes them so the policy can act without external intervention. A composite reward combines answer correctness with tool-usage regularization.

Rollout Modes

  • Rollouts sample between controlled and uncontrolled modes; in the controlled mode, a controller triggers a signal when it detects abnormal retrieval, while the other mode conditions only on task and state.The policy conditions its next action on the current reasoning state and control signal when one is triggered.Source: §B.3 · pp. 16–17

Update Modes

  • An annealed control-forcing curriculum shifts training from frequent control signals to no control, improving early stability while transferring learned behavior to the no-control setting.Rollouts are generated by the current policy under each observation regime, with on-policy updates; the final training stage removes control entirely.Source: §B.3 · p. 17

Reward Design

  • The composite outcome-driven reward combines F1-based answer correctness with penalties for improper tool use and control non-compliance.Violations include malformed tool inputs and failure to follow control messages; penalties are scaled and capped, while a format floor gives valid outputs non-zero reward.Source: §B.3 · p. 17

C Dataset Overview

DEEPCONTROL is evaluated on seven question-answering datasets spanning general and multi-hop QA. Dataset splits come from FlashRAG’s curated dataset collection.

C Dataset Overview

  • Evaluation covers NQ, TriviaQA, and PopQA for general QA, plus HotpotQA, 2WikiMultiHopQA, Musique, and Bamboogle for multi-hop QA; splits come from FlashRAG.NQ, TriviaQA, and PopQA are the general-QA datasets; the other four require multi-hop reasoning or compositional questions.Source: C Dataset Overview · p. 17
  • NQ has 79,168 training and 3,610 test samples, TriviaQA uses an 11,313-example test set, and PopQA contains 14,267 entity-centric triples.NQ pairs real Google search queries with human-annotated Wikipedia answers; TriviaQA is a reading-comprehension benchmark, while PopQA measures parametric knowledge coverage on long-tail entities.Source: C Dataset Overview · p. 17
  • HotpotQA has 90,447 training and 7,405 development samples; 2WikiMultiHopQA uses a 12,576-example development split, Musique 2,417, and Bamboogle 125.HotpotQA requires reasoning across multiple Wikipedia paragraphs; 2WikiMultiHopQA combines structured and unstructured Wikipedia information, Musique forms 2–4-hop problems, and Bamboogle contains two-hop questions selected for search-engine errors.Source: C Dataset Overview · p. 17

D Implementation Details

DeepControl uses hierarchical selective expansion to regulate how much retrieved evidence reaches the agent, with training controls designed to teach acquisition behavior that the final policy can use without external control. The implementation uses shared retrieval and training settings across tasks, while an exploratory richer corpus variant is excluded from the main results to preserve comparability.

Hierarchical Corpus Construction

  • At inference, retrieval first exposes passage titles and first-sentence summaries; an expand action reveals the full passage in a two-level extractive hierarchy.The main experiments use the 2018 Wikipedia dump, matching prior search-augmented RL baselines.Source: D Implementation Details · p. 18
  • The exploratory three-level hierarchy uses generated document and section summaries over Wikipedia pages, but is excluded from main results because its corpus differs from prior baselines.Retrieved text is also cleaned when it contains formatting artifacts.Source: D Implementation Details · p. 18

Control Hyperparameters

  • Utility combines novelty and answer-score improvement to control stopping and continuation, using shared thresholds across tasks and showing insensitivity to moderate parameter variations.Passages are encoded with E5 after truncation to 512 tokens; novelty uses kNN with knn = 5, while continuation also requires a low gold-answer score.Source: D Implementation Details · p. 18

Reward Hyperparameters

  • Reward settings provide validly formatted outputs a 0.1 floor and cap per-violation tool-use penalties at 0.4, with training robust to moderate variations.The per-violation penalty is 0.2; the format floor is intended to stabilize early training without overshadowing answer correctness.Source: D Implementation Details · p. 18

Training Setup

  • DEEPCONTROL trains with PPO on merged NQ and HotpotQA data, annealing control probability from 0.9 to 0 across five epochs and evaluating Exact Match on seven datasets.GRPO led to training collapse after a few dozen optimization steps; the method retrieves five candidate summaries but permits at most three expansions, matching baselines’ three-passage evidence budget (Lin et al., 2023).Source: D Implementation Details · p. 18

Training Cost

  • For 1,000 steps, training wall-clock time rises from 3.61 to 4.66 days, and average rollout time from 5.2 to 6.7 minutes; inference adds no runtime overhead.The added cost occurs only during training, mainly from utility estimation and controller construction.Source: D Implementation Details · p. 19

E Additional Results

This section presents additional results and analyses of adaptive information control, including its robustness, stopping behavior, and failure modes. It also provides sample search-augmented answers and examples involving suboptimal search behavior and retrieved material about Corliss Archer adaptations.

DEEPCONTROL vs. Vanilla PPO; PPO vs. GRPO Under Information Control

  • DEEPCONTROL improves performance by 8.3% over vanilla PPO; under control annealing, PPO is more stable and achieves higher final no-control performance than GRPO.GRPO improves faster early in training, but later reward degradation is accompanied by KL and policy-gradient-loss spikes and a sharp decline in valid tool actions.Source: E Additional Results · pp. 19–20, Table 3 · p. 20

Alternative Utility Formulations

  • The default additive utility performs best overall, while binary-effectiveness and multiplicative alternatives remain competitive but are less effective.The comparison keeps other training, control, and reward settings fixed; results favor continuous effectiveness shaping and an additive combination.Source: Table 4 · p. 21, E Additional Results · p. 21

Gold-answer-free Utility Estimation

  • A gold-answer-free decision-impact variant still outperforms Search-R1 baselines, showing that the gains do not solely depend on gold-answer access.The variant estimates utility from evidence-induced changes in the model’s answer distribution and uncertainty.Source: Table 5 · p. 21, E Additional Results · p. 21

No-Control Behavior and Stopping Quality

  • Under no-control inference, DEEPCONTROL searches less on NQ and more on HotpotQA; forced continuation reduces accuracy, while removing the final search step also substantially reduces it.These interventions indicate that the policy generally does not stop prematurely and that its final retrieval is often useful rather than redundant.Source: Table 6 · p. 22, Table 7 · p. 22, E Additional Results · p. 23

Control-hyperparameter and Reward-hyperparameter Sensitivity Analysis

  • Performance remains stable under moderate control-hyperparameter perturbations, while reward sensitivity is greatest for the malformed-action penalty.Intermediate control settings offer the best under-search/over-search tradeoff; retrieval bonus and format floor have milder effects, and setting λceil = 1.0 degrades performance.Source: E Additional Results · p. 22

Schedule Sensitivity Analysis

  • The default annealing schedule achieves the best final no-control performance, and later checkpoints show higher early-step utility as external control is reduced.The tested stages use p = 0.9, 0.5, 0.2, and 0; faster annealing weakens early guidance, while slower annealing delays policy internalization.Source: E Additional Results · p. 22

Control vs. Reward Shaping; Effect of Hierarchical Evidence Construction

  • Information control remains beneficial with outcome-only rewards, whereas hierarchical evidence construction without adaptive control underperforms Search-R1.The findings attribute the gains to adaptive information control rather than reward shaping alone, and indicate that control guidance helps agents learn when to expand evidence.Source: Table 8 · p. 23, Table 9 · p. 23, E Additional Results · p. 23

Robustness Analysis; Error Analysis; Example Outputs

  • DEEPCONTROL also outperforms Search-R1 with BM25; error analysis finds fewer major behavioral failures overall, though reasoning errors remain the largest residual source.The paper notes that current experiments focus on Qwen models; examples illustrate interleaved retrieval and selective expansion guided by continuation and termination controls.Source: E Additional Results · pp. 24–25

Suboptimal Search Behavior (Search-R1-qwen2.5-3b-it-grpo on NQ)

  • The cited Scandal exchange says the show was in its seventh season, that the current season’s last episode premiered January 26, 2017, and that the next episode was coming out on January 26, 2017.Source: E Additional Results · p. 27

Suboptimal Search Behavior (Search-R1-qwen2.5-3b-it-grpo on HotpotQA)

  • In this HotpotQA exchange, a search-result passage identifies Simon Fuller as David Beckham’s recruiter, but the final response gives “2003 -2009” without displayed passage evidence establishing his Manchester United tenure.Source: E Additional Results · pp. 28–29

Output Format Collapse under Large Top-K Retrieval (Search-R1-qwen2.5-3b-it-grpo on HotpotQA)

  • The appendix labels a HotpotQA example as output-format collapse under large top-k retrieval and begins with a question about the government position held by the actress who portrayed Corliss Archer.The supplied excerpt shows the prompt and the start of the search result, but not the assistant’s collapsed output.Source: E Additional Results · p. 29

E Additional Results

  • The excerpts also note a 1949 sequel, A Kiss for Corliss, and television adaptations of Meet Corliss Archer from 1951–1952 and 1954–1955.The syndicated television show starred Ann Baker and Mary Brian.Source: E Additional Results · p. 30
  • The excerpts describe the 1945 film as starring Shirley Temple as Corliss Archer and adapting F. Hugh Herbert’s Broadway play and short stories.The play was written by Herbert, who also wrote the film; the film was directed by Richard Wallace.Source: E Additional Results · p. 30
  • The retrieved response says the woman who portrayed Corliss Archer held the position of Secretary of State, but the supplied excerpts do not support that link.The snippets identify Shirley Temple as Corliss Archer in the 1945 film, but provide no evidence about her holding a government position.Source: E Additional Results · pp. 30–31
  • Live performances aired in 1951 and 1952, followed by a 1954–1955 syndicated show starring Ann Baker and Mary Brian.Episodes occasionally cut to comic-book-style drawings with announcer commentary to illustrate the story, several times per episode.Source: E Additional Results · p. 34

Prompt: Search-Augmented Reasoning

This section presents a structured search-augmented reasoning prompt in which the assistant iteratively retrieves and selectively expands documents before producing a concise, normalized answer. Its examples illustrate control-guided decisions about continuing or stopping search and answering from retrieved evidence.

Prompt: Search-Augmented Reasoning

  • Control messages can direct the agent to expand retrieved documents, continue searching, avoid expansion, or stop searching.The protocol says control messages may appear anywhere in the conversation and must be followed.Source: Prompt: Search-Augmented Reasoning · pp. 35–36
  • Final answers must exactly match the canonical short answer: output the shortest possible answer span, without explanations, appositives, parentheses, extra words, punctuation, or formatting, and use case-sensitive matching.Examples give “seven” for the episode count and “February 20, 2017” for the release date.Source: Prompt: Search-Augmented Reasoning · p. 35
  • Each round requires reasoning, searching for snippets, and optionally expanding selected documents into full text before continuing or answering.The prompt permits multiple rounds and specifies an answering-round and a continuing-round sequence.Source: Prompt: Search-Augmented Reasoning · pp. 35–36
  • In the India-election example, the agent expands a relevant search result and answers “1951–52” after finding the first Lok Sabha elections ran from October 25, 1951, to February 21, 1952.The example shows a search followed by expansion of document 17180261 and a concise final answer.Source: Prompt: Search-Augmented Reasoning · p. 37
  • In the Tamar Ziskind example, an initial search and expansion fail to identify the relevant site, so the agent searches more specifically for Tamar Ziskind and Haifa.The supplied example ends after expanding the Haifa result, without showing the final answer.Source: Prompt: Search-Augmented Reasoning · pp. 38–40
  • The Haifa example illustrates repeated searching and document expansion before answering that the Bahá'í World Centre is in Tamar Ziskind’s hometown.After initial results did not show the UNESCO information, the assistant expanded the Haifa document and then gave the short answer.Source: Prompt: Search-Augmented Reasoning · pp. 40–41
  • The protocol requires reasoning after new search results or information, uses <search> for retrieval and <expand> for full text, and permits repeated rounds until answering.An answering round ends with <answer>; a continuing round ends after reasoning, expansion, and information without a final answer.Source: Prompt: Search-Augmented Reasoning · pp. 41–42
  • Answer normalization requires the exact canonical short answer, with no explanation or extra formatting and case-sensitive matching.The prompt directs the assistant to use the most common standalone name form and shortest answer span.Source: Prompt: Search-Augmented Reasoning · pp. 41–42
  • In the Reading F.C. example, a stop-search control prevents further retrieval, and the assistant answers “Xiu Li Dai” from the ownership information already available.The expanded information states that 75% is owned by Renhe Sports Management Ltd, itself 100% owned by Xiu Li Dai and Yongge Dai.Source: Prompt: Search-Augmented Reasoning · pp. 43–44
  • After initial results failed to identify KY 46, the search was redirected to that route and county, then its retrieved document was expanded.The initial results included routes 47 and 43; the targeted search returned KY 46 in Nelson County, and the control prompt requested expansion of its document.Source: Prompt: Search-Augmented Reasoning · pp. 45–46
  • Because the expanded route information lacked population data, the search separately looked up Nelson County’s population and stopped when instructed.The returned population result reported 43,437 in the 2010 census; the final response followed the stop-search instruction.Source: Prompt: Search-Augmented Reasoning · pp. 46–47
  • Nelson County is the answer: Kentucky Route 46 runs through it, and its 2010 population of 43,437 exceeds 43,000.The final response identifies Nelson County as meeting both conditions.Source: Prompt: Search-Augmented Reasoning · pp. 46–47

Related papers

Loading 2602.01672v2…