PaperlayerView exact revision on arXiv

Source-linked AI summary

PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces

M. Yigit Turali, Shreyas Rajesh, M. Efe Lorasdagi, Chenda Duan, Kartik Sharma, Vwani Roychowdhury

arXiv:2610.15208v1cs.CLcs.AIcs.IR

TL;DR

Deep research requires iterative searches and reasoning, with findings guiding what to investigate next and explicit state helping agents manage extended investigations. PANINI-DR coordinates subquestion research and verification through a shared workspace. It achieves the strongest average performance across three benchmarks, and leads when correct answers must be recoverable from saved evidence.

  • Problem

    Deep research requires connecting evidence across sources and carrying findings across iterative searches, while dependencies allow errors in intermediate findings to propagate.

    Source: §1 · pp. 1–2

  • Method

    PANINI-DR coordinates researchers and a separate verifier through a shared workspace that records subquestions, dependencies, findings, supporting evidence, and review outcomes.

    Source: §1 · p. 2

  • Results

    PANINI-DR achieves the strongest average answer quality across three benchmarks, with particularly large gains on harder BrowseComp-Plus questions.

    Source: §1 · p. 3, §6 · p. 9

  • Takeaways & Limitations

    Accounting for whether answers are recoverable from saved evidence changes harness comparisons and further strengthens PANINI-DR’s advantage.

    Source: §1 · p. 3

    The primary comparisons use one research backbone and fixed benchmark corpora, leaving generalization to other models and open-web research untested.

    Source: §6 · p. 9

Abstract

from arXiv · show

Deep research requires connecting multiple steps of search, reasoning, and computation, with intermediate findings guiding what to investigate next. As investigations grow, agents must maintain research state across model calls and check findings before errors propagate into subsequent work. We introduce PANINI-DR, a multi-agent harness that separates coordination of the overall investigation from the research and verification of individual subquestions. A shared workspace, built on the Generative Semantic Workspace representation, records subquestions, their dependencies, findings, and supporting evidence. Researchers investigate individual subquestions, while a separate verification agent checks proposed answers against those subquestions and their evidence. Verification feedback guides further investigation, and recorded findings inform subsequent assignments. We evaluate PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven state-of-the-art deep-research harnesses, all using a shared language-model backbone and frozen corpora. PANINI-DR achieves the strongest average performance across the three benchmarks. On BrowseComp-Plus, it achieves 73.3% accuracy, outperforming the next-best harness by 16.0 points. Its advantage is strongest on hard questions, where it answers nearly twice as many as the next-best harness. Ablations further show that structured planning and persistent state improve performance, as does checking each finding against its cited evidence before it is reused. Because a model can answer correctly from memory alone, we also analyze the evidence retained by each harness. PANINI-DR achieves the highest answer recoverability from saved evidence on all three benchmarks, while complementary claim-support audits show stronger evidence-backed answers. Accounting for this evidence changes how harnesses compare and further strengthens PANINI-DR's advantage.

1 INTRODUCTION

Deep research requires iterative search and reasoning because relevant sources may be unknown at the outset and findings can shape what to investigate next. PANINI-DR combines explicit research state with verification of intermediate findings, and achieves the strongest average performance across the evaluated benchmarks.

1 INTRODUCTION

  • Deep research requires connecting evidence across sources through successive searches and reasoning, since relevant documents may not be apparent in a single retrieve-then-answer step.Intermediate findings can guide later investigation and computation.Source: §1 · pp. 1–2
  • Implicit state in a growing interaction history makes earlier findings harder to use, while dependencies between subquestions let errors propagate into later work.Verifying findings against their evidence and intended subquestions offers an opportunity to catch errors before they guide subsequent research.Source: §1 · p. 2
  • PANINI-DR separates coordination from subquestion research and verification, recording dependencies, findings, evidence, and review outcomes in a shared workspace.Researchers investigate individual subquestions; verification feedback guides further investigation, and recorded findings inform subsequent assignments.Source: §1 · p. 2
  • PANINI-DR achieves the strongest average performance across benchmarks, with its largest gains on hard questions and stronger comparisons when answers must be recoverable from saved evidence.The paper also reports that accounting for evidence changes harness comparisons and further strengthens PANINI-DR’s advantage.Source: §1 · p. 3

2 THE PANINI-DR HARNESS

PANINI-DR coordinates research and verification through a shared workspace that records subquestions, dependencies, findings, and supporting evidence. Its Orchestrator uses this evolving state to direct research, admit checked findings, and produce an answer.

2.1 CONSTRUCTING THE RESEARCH WORKSPACE

  • The Orchestrator initializes a GSW plan of subquestions and dependencies, then pairs it with per-node records of findings, evidence, and review status.Structural checks and bounded repairs address invalid references, answer targets, circular dependencies, and hints inconsistent with plan dependencies.Source: §2.1 · pp. 3–4
  • The dependency graph links subquestions to later questions or the final answer, while blanks carry retrieval hints and verification metadata such as expected answer type and required claim.The workspace begins with unanswered blanks and is updated as findings are recorded.Source: §2.1 · pp. 3–4

2.2 RESEARCHING A SUBQUESTION

  • Researchers first try one retrieval round with available findings substituted into the hint, then use iterative ReAct search and reasoning when the needed context or a valid candidate is missing.Candidates, sources, searches, and verification feedback can be retained for later assignments, allowing research to build on earlier attempts.Source: §2.2 · p. 5

2.3 VERIFYING FINDINGS AND UPDATING THE WORKSPACE

  • Verification combines type and evidence-attachment checks with assessment of whether cited evidence supports the candidate for its subquestion and distinguishes plausible alternatives.Failed checks return feedback for revision or further evidence gathering; admitted findings fill workspace blanks and retain their evidence and review status.Source: §2.3 · p. 5

2.4 ANSWERING THE QUERY

  • Admitted findings supply context for later research and inputs for planned computations, after which the Orchestrator submits the answer when its target is resolved.Researchers can suggest missing intermediate subquestions for plan revision; if the investigation ends without an answer, a final LLM call may synthesize one from collected evidence without further search.Source: §2.4 · pp. 5–6

3 EXPERIMENTAL SETUP

The evaluation covers three complementary benchmarks and compares PANINI-DR with seven harnesses under a shared model backbone and frozen benchmark corpora. It measures answer quality, recoverability from saved evidence, and estimated research computation.

3.1 DATASETS

  • BrowseComp-Plus tests entity identification from detailed clues, FRAMES tests fact combination across sources, and MoNaCo tests collection-valued answers.The evaluation uses the full 824-question FRAMES and 1,315-question MoNaCo sets, plus 150 questions across the BCP easy, medium, and hard groups.Source: §3.1 · p. 6

3.2 EVALUATION METRICS

  • Answer quality uses official accuracy on BCP and FRAMES and mean item-level F1 on MoNaCo, while observed support credits correct content only when recoverable from saved evidence.MoNaCo observed-support F1 is recomputed using only judge-credited items that pass the support rule.Source: §3.2 · p. 6

3.3 BASELINES

  • The seven baselines include retrieval-program, trajectory-compression, evolving-report, evidence-tracking, memory, and CodeAct search-and-tool approaches.The evaluation uses released implementations where available and reimplements Omni-Decision and HyMem from their published specifications.Source: §3.3 · pp. 6–7

3.4 IMPLEMENTATION DETAILS

  • All systems use gpt-oss-120b and, within each benchmark, the same frozen corpus and retrieval service; research allowances and retrieval breadth vary by system.PANINI-DR assignments allow 20 ReAct turns and up to five retries after failed checks.Source: §3.4 · p. 7

4 RESULTS AND DISCUSSION

PANINI-DR leads the three-benchmark average, with its advantage largest on hard questions, and also leads when answer credit requires recoverability from saved evidence. Ablations report benefits from structured planning and state and from verification before findings are reused, alongside higher research-computation costs.

4.1 ANSWER QUALITY

  • PANINI-DR has the best average across the three benchmarks, nearly doubles SLEUTH’s accuracy on hard BCP questions, and retains a 16-point BCP gain over SLEUTH after removing questions SLEUTH refuses.The advantage is largest on the hardest questions; the comparison is consistent across three runs per harness.Source: §4.1 · p. 7

4.2 EVIDENCE QUALITY

  • PANINI-DR leads observed support on all three benchmarks, with a wider margin over SLEUTH than on official scores, and remains ahead under a model-based claim-support audit.Its saved evidence also yields the highest answer quality in a common-reader test on all three benchmarks, whether the reader may abstain or is told an answer is present.Source: §4.2 · pp. 7–8

4.3 ABLATIONS

  • 58.0% versus 41.3% BCP accuracy results when adding a plan and persistent state; with matched final synthesis, accuracy is 73.3% versus 56.7%.New entities appear in later searches 77.4% of the time with planning and state, compared with 32.1% without them.Source: §4.3 · p. 8
  • The full verification policy with up to five retries adds 16.0 points to BCP accuracy and observed support versus removing the Verifier and citation and evidence-attachment checks.Assignment and answer-type checks remain in both conditions, and gains also hold on matched FRAMES and MoNaCo subsets.Source: §4.3 · p. 8

4.4 COMPUTATION AND RESEARCH ALLOWANCES

  • PANINI-DR’s estimated computation is 7.3× SLEUTH’s despite sending 19.2× its prompt tokens; 66.5% of PANINI-DR input tokens came from cache in recorded BCP runs.Eighty Researcher turns raise BCP accuracy from 73.3% to 76.0% but use 2.8× the default’s prompt tokens, so the authors retain the 20-turn default.Source: §4.4 · pp. 8–9

5 RELATED WORK

Prior work addresses explicit reasoning, plans, persistent research state, intermediate verification, and source attribution. PANINI-DR connects these strands through an execution protocol linking typed assignments, subquestion-scoped review, and admitted values used in later work.

Research state and executable plans

  • PANINI-DR’s contribution is an execution protocol connecting typed research assignments, subquestion-scoped review, and admitted values used by subsequent work.It builds on the GSW representation rather than contributing a dependency graph alone.Source: §5 · p. 9

Verification during research

  • PANINI-DR ties verification to persistent research subquestions: a Verifier checks candidates against verification claims and cited evidence, and the harness controls admission while retaining review status.Prior work includes process supervision, verifier-guided search, draft checking, retrieval critiques, and fine-grained model-based evaluation.Source: §5 · p. 9

Attribution and support

  • Attribution work evaluates source support for generated claims, while observed support measures deterministic answer recoverability from saved evidence.The paper discusses extended related work in Appendix E.Source: §5 · p. 9

6 LIMITATIONS AND CONCLUDING REMARKS

PANINI-DR combines explicit research state with verification of intermediate findings and achieves the strongest average answer quality across three deep-research benchmarks. Its evaluation and verification remain bounded by the planner, available evidence, fixed benchmark setting, and substantial computation.

6 LIMITATIONS AND CONCLUDING REMARKS

  • PANINI-DR achieves the strongest average answer quality across three deep-research benchmarks, with particularly large gains on harder BrowseComp-Plus questions.Observed-support evaluation further widens its advantage when correct answer content must be recoverable from retained evidence.Source: §6 · p. 9
  • Ablations support the value of both structured research state and intermediate verification.The harness is built around maintaining explicit research state and verifying findings before they guide subsequent work.Source: §6 · p. 9
  • Verification depends on planner-identified subquestions, research candidates, and evidence available to the Verifier; model judgments can be wrong and checks cannot guarantee complete plan coverage.Observed support measures recoverability from saved evidence, not whether every claim is supported; complementary claim-support audits also rely on model judgments.Source: §6 · p. 9
  • Primary comparisons use one research backbone and fixed benchmark corpora, leaving generalization to other models and open-web research untested.The harness requires substantial computation, and increasing the verification allowance does not consistently improve answer quality.Source: §6 · p. 9

ETHICS STATEMENT

The study evaluates its proposed methods on existing benchmark datasets and reports no additional ethical concerns specific to the study beyond those associated with language models and their applications.

ETHICS STATEMENT

  • The study evaluates its proposed methods on existing benchmark datasets.Source: ETHICS STATEMENT · p. 10
  • The authors identify no additional ethical concerns specific to this study beyond those associated with language models and their applications.Source: ETHICS STATEMENT · p. 10

AI USE STATEMENT

Language models served as system components and graders, and AI assistance was used to edit the draft and catch grammatical errors. The authors take responsibility for the paper’s contents.

AI USE STATEMENT

A.1 SHARED EVALUATION SETUP

This section specifies shared benchmark and grading protocols, tests whether the BCP ranking survives a grader change, and illustrates how recorded investigations turn clues into structured research plans.

A.1 SHARED EVALUATION SETUP

  • The harness comparison shares a backbone, corpus, and retrieval service, but does not equalize computation.The setup describes benchmark-specific difficulty groups and question formats; Table 3 lists baseline protocols.Source: §A.1 · p. 17, Table 3 · p. 17

A.2 ANSWER GENERATION AND GRADING

  • BCP and FRAMES use majority voting over three judgments, while MoNaCo uses the highest-F1 complete judgment for scoring and support accounting.Blank FRAMES answers score zero without a grader call; missing judgments score zero, and unreadable judgments do not count as correct.Source: §A.2 · p. 17
  • The sampling control selects each question’s most frequent non-empty answer across repeated SLEUTH runs.This answer-selection vote is separate from the three grading judgments; accuracy and Obs. use the selected run’s verdict and saved evidence.Source: §A.2 · p. 17

A.3 SENSITIVITY TO THE GRADING MODEL

  • Table 4 reports grader sensitivity on BCP CF-150 (150 questions), comparing gpt-5.6-luna (primary) and gpt-5.6-terra on the same answers, official prompt, and 3-pass majority.Scores are correct-answer counts with accuracy (%) in parentheses; ∆ is in percentage points. Agreement is the percentage of matching verdicts, and κ is Cohen’s kappa.Source: Table 4 · p. 18

B IMPLEMENTATION THROUGH RECORDED INVESTIGATIONS

B.1 FROM THE QUESTION TO A RESEARCH PLAN

  • For the film question, the plan links the target to its credits and biographies, with deterministic checks for runtime, third-film status, and birth-year bounds.The cap-5 execution contains 9 given anchors, 13 relations, 12 unknown-value plan nodes, and 4 deterministic constraints.Source: §B.1 · p. 18
  • The plan representation records anchors, nodes, dependencies, computations, and the answer target; validation checks references and rejects dependency cycles.Advisory tags may flag omissions or ambiguity, but do not block dispatch or change the plan.Source: §B.1 · p. 18
  • Assignments follow the unresolved frontier whose dependencies are already given or admitted, while prerequisite research outside that frontier remains allowed.In this example, the plan remains unchanged through three dispatches requesting 4, 3, and 1 research nodes.Source: §B.1 · p. 19

B.2 BRAZUCA: FINDINGS, FEEDBACK, AND EVIDENCE COMPLETION

The walkthrough shows how findings are researched, checked, and revised across several paths, then describes tests of evidence recoverability and support.

B.2 BRAZUCA: FINDINGS, FEEDBACK, AND EVIDENCE COMPLETION

  • Researchers try typed retrieval first, continue in ReAct when it fails, and submit cited findings for checks of support, satisfaction, and uniqueness.ReAct relation attachments are checked for scope and identifiers before the Verifier assesses whether the finding satisfies the assigned claim.Source: §B.2 · p. 19
  • Verification feedback refines an ambiguous editor identity and prompts added credits; three passages then secure endorsement of the Brazuca target.A non-unique surname is rejected before the Researcher resubmits the full name; the target is endorsed after the evidence establishes its connections.Source: §B.2 · pp. 19–20

B.3 WHEN RESEARCH TAKES A DIFFERENT PATH

  • Across alternate paths, the harness handles list findings and bounded retries, can add missing subquestions, but revision may leave affected resolved findings stale.Warnings do not automatically schedule retries, and a revision can leave an incorrect target resolved.Source: §B.3 · pp. 20–21

B.4 FORT SANTIAGO: ANSWERING FROM THE RETAINED RESEARCH

  • For Fort Santiago, unresolved relation bindings block admission despite a recoverable name; the Finalizer can answer from retained sources without changing bindings or verification.The Finalizer returns Fort Santiago after the research loop stops, and all 3 grading judgments accept it.Source: §B.4 · p. 21

C EVIDENCE QUALITY AND VALIDATION

  • Obs. measures recoverability from an assembled record of saved tool outputs, validated quotations, and recorded inputs, rather than confirmed receipt of every passage.For BCP and FRAMES, the scoring procedure combines official answer credit with evidence matching; MoNaCo uses item-level credit and support.Source: C EVIDENCE QUALITY AND VALIDATION · p. 22

C.1 CHECKING THE SCORER AND EVIDENCE RECORD

  • Removing identifier-reconstructed bodies leaves all 7 assessable systems’ BCP Obs. scores unchanged; PANINI-DR also remains highest on all 3 benchmarks under Obs.+.Its BCP score is 70.7% versus SLEUTH’s 54.0%, a paired gain of 16.7 points.Source: §C.1 · p. 23

C.2 MODEL-BASED GROUNDED-ONLY AUDIT

  • After unsupported claims are filtered, PANINI-DR leads both baselines on all 3 benchmarks.The answers are rebuilt from claims judged supported in context and re-graded under each benchmark’s official protocol.Source: Table 5 · p. 23, §C.2 · p. 23

C.3 ANSWERING FROM EACH SYSTEM’S EVIDENCE

  • PANINI-DR’s evidence yields the highest answer quality on all 3 benchmarks under both reader prompts.With abstention allowed, it scores 62.7% versus 55.3% on BCP, 69.4% versus 62.5% on FRAMES, and 0.547 versus 0.515 on MoNaCo.Source: Table 6 · p. 24, §C.3 · p. 24

C.4 CHECKING COMPUTED ANSWERS

The calculation audit supplements text matching by replaying source-validated inputs; follow-on analyses examine source coverage, plan revision and recovery, and answering with supplied sources.

C.4 CHECKING COMPUTED ANSWERS

  • Obs-D supplements text matching by checking whether saved inputs are recoverable from cited documents and replaying recorded operations to reproduce the graded answer.A computed answer may not appear in a source passage; in the Camp David example, replay recovers 51 from the birth and signing dates.Source: §C.4 · p. 24
  • The audit adds credit only for official credit not already retained by Obs, with FRAMES additions binary and MoNaCo retaining partial credit for the derived item.All questions remain in the denominator.Source: §C.4 · p. 24, Table 7 · p. 25
  • Replay adds credit on 51 FRAMES questions, raising scores from 66.1% to 72.3%, and on 55 MoNaCo questions, raising scores from 0.458 to 0.498.Restricting additions to calculations with accepted Verifier records for every retrieved input leaves the gains unchanged.Source: §C.4 · p. 24, Table 7 · p. 25

D.1.1 FINDING SOURCES AND CONNECTING FINDINGS

  • PANINI-DR leads SLEUTH on source recall, coverage, and precision on every benchmark; on BCP, it records every annotated page on 121 questions versus 46.Page presence includes snippets and recorded typed-resolver and post-loop grounding inputs, not just full-page reads.Source: §D.1.1 · p. 25, Table 8 · p. 25

D.1.2 REVISING THE PLAN AND RECOVERING AN ANSWER

  • Accepted plan revisions add 59 BCP and 162 FRAMES subquestions, with 30 and 82, respectively, holding a value at the end.These counts describe added research tasks and their completion, not an accuracy effect from turning revisions on or off.Source: §D.1.2 · p. 25
  • On BCP, Finalizer synthesis or stored-candidate fallback supplies 40 of 110 correct answers, including 38 that pass Obs.These recovery paths apply when the target remains unfinished.Source: §D.1.2 · p. 25, Table 9 · p. 26

D.1.3 ANSWERING WITH ANNOTATED SOURCES

  • With annotated documents supplied and no search, planning, or verification loop, gpt-oss-120b reaches 93.3% BCP accuracy, including 45/50 hard questions versus PANINI-DR’s 26/50.This is a diagnostic rather than a cost-matched comparison or universal accuracy ceiling; source-packaging procedures differ across benchmarks.Source: §D.1.3 · p. 26, Table 10 · p. 26
  • The supplied-source reference answers 14 of 17 BCP questions where PANINI-DR found all annotated documents but answered incorrectly.In one example, research reached the sources but connected the target to the wrong artist’s death.Source: §D.1.3 · p. 26

D.2.1 STRUCTURED PLANNING AND PERSISTENT STATE

These analyses test structured planning and persistent state, verification before reuse, bounded review, token use, repeated sampling, and longer research assignments. Together, they show gains from planning/state and verification while qualifying the costs and limits of the tested policies.

D.2.1 STRUCTURED PLANNING AND PERSISTENT STATE

  • 16.7 points: ON outperforms OFF in accuracy; the gain persists with common recovery, while the tested intervention is a planning-and-state package.ON has 37 answers become correct and 12 become incorrect; common recovery adds 23 correct answers to each arm.Source: §D.2.1 · pp. 26–27
  • 77.4% of eligible recorded entities are reused in later searches on ON, versus 32.1% on OFF; sampled checks validate many matches but not their causal role in accuracy.The contextual check labels 37 of 40 ON excerpts and 28 of 40 OFF excerpts as genuine reuse.Source: §D.2.1 · p. 27

D.2.2 VERIFICATION BEFORE REUSE

  • 16.0 points: full verification improves both official and observed-support scores on all three matched sets.On BCP, median prompt use rises from 0.453 M to 1.042 M, including research prompted by feedback.Source: §D.2.2 · p. 28

D.2.3 RESPONDING TO VERIFICATION FEEDBACK

  • 73.3% accuracy: cap 5 exceeds cap 10’s 66.0%, while cap 2 is close; more review opportunities do not consistently improve answers.At cap 5, 77 of 101 endorsed proposals following rejection change the value or citation set.Source: §D.2.3 · p. 28

D.3.1 RECORDED TOKENS AND CACHE REUSE

  • 85–89% of prompt tokens go to the Researcher’s ReAct stage and Reader, versus 4–6% for the Orchestrator and 2–3% for the Verifier; cache reuse narrows the estimated computation gap with SLEUTH.PANINI-DR submits 19.2 times SLEUTH’s prompt tokens, but reuses 66.5% through the prefix cache; estimated model-core computation is 7.3 times SLEUTH’s.Source: §D.3.1 · pp. 28–29

D.3.2 REPEATED SAMPLING AT COMPARABLE COMPUTATION

  • 64.7% accuracy: at eight runs per question, SLEUTH uses 1.12 times PANINI-DR’s model-core computation but trails its 73.3% accuracy by 8.7 points.Under the tested voting policy, accuracy peaks at seven runs; even an oracle choosing any correct run would answer only 70.0% of questions.Source: §D.3.2 · p. 29

D.3.3 EXTENDING INDIVIDUAL INVESTIGATIONS

  • 76.0% accuracy: extending Researcher allowances from 20 to 80 turns raises accuracy from 73.3%, but the longest allowance uses 2.8 times the prompt tokens for 2.7 additional points.Longer runs also complete support for existing candidates: 24 and 26 questions shift from correct recovery answers to correct answers with resolved workspace targets in the 50- and 80-turn runs.Source: §D.3.3 · p. 30
  • SLEUTH’s allowance sweep shows no sustained accuracy gain as mean prompt use rises from 74k to 111k; turn-policy comparisons are not computation-matched.From 20 to 80 turns, budget exhaustion falls from 7 to 1 question, while input-length refusals rise from 19 to 26.Source: §D.3.3 · pp. 30–31

D.4 RUN STABILITY, PAIRED COMPARISONS, AND ANSWER STRUCTURE

The discussion reports run stability and benchmark-specific qualifications, then situates PANINI-DR among related work on reasoning structure, persistent memory, and verification. It also distinguishes the paper’s protocol from prior approaches and notes evaluation and implementation boundaries.

D.4 RUN STABILITY, PAIRED COMPARISONS, AND ANSWER STRUCTURE

  • 72.7% mean BCP accuracy across three runs exceeds SLEUTH’s and Open Deep Search’s 58.0%; PANINI-DR’s lowest run also exceeds every baseline run in this comparison.On FRAMES, the accuracy advantage over SLEUTH has a paired interval including zero; on MoNaCo, PANINI-DR trails four harnesses overall and its collection recall is lower than HyMem’s.Source: §D.4 · p. 31
  • On the 131 BCP questions remaining after removing the same 19 refusal-affected questions from both systems, PANINI-DR’s accuracy advantage over SLEUTH is 9.2 points and its observed-support advantage is 9.9 points.The analysis removes recorded refusals rather than predicting performance under a different context window.Source: §D.4 · p. 31

D.5 BASELINE ADAPTATIONS

  • The comparison includes paper-based reimplementations as well as released-code adaptations; Re-TRAC’s 76 empty outputs and seven abstentions score zero on BCP under the tested entry point.SLEUTH, Omni-Decision, and HyMem had no available code when experiments were established, so the authors followed published specifications.Source: §D.5 · pp. 31, 33

E.1 REASONING STRUCTURE, PLANS, AND SEMANTIC WORKSPACES

  • Prior work structures generated reasoning, follow-up retrieval, computations, or corpus evidence; PANINI-DR instead carries admitted, evidence-linked intermediate values across research calls.Its subquestions persist as assignments and claims to check, while evidence and review outcomes stay attached to admitted values.Source: §E.1 · pp. 33–34

E.2 DEEP-RESEARCH HARNESSES AND PERSISTENT MEMORY

  • Deep-research harnesses retain information in different forms, including rollout summaries, evolving reports, and tiered memories; PANINI-DR uses subquestion-bound values and review status to guide later assignments.The discussion distinguishes this execution protocol from answer-completion policies such as recovery from retained articles.Source: §E.2 · pp. 34–35

E.3 VERIFICATION, FEEDBACK, AND CORRECTIVE RETRIEVAL

  • Prior verification work uses checking to select, revise, or support retrieval; PANINI-DR checks intermediate values, returns negative verdicts for further research, and separately applies an admission policy.Unlike final-answer sufficiency approaches, its negative verdict returns a proposal to the Researcher under a shared review cap.Source: §E.3 · p. 36

E.4 ATTRIBUTION AND EVIDENCE-BASED EVALUATION

The section distinguishes evidence availability and recorded computation from question-conditioned attribution, and describes provenance as a possible training signal distinct from PANINI-DR’s training-free measurement and control.

E.4 ATTRIBUTION AND EVIDENCE-BASED EVALUATION

  • Observed support measures whether an answer is recoverable from the execution evidence record while retaining official answer credit; Obs-D additionally checks cited inputs and replays recorded calculations.These measures complement, rather than replace, question-conditioned attribution.Source: §E.4 · p. 37
  • STAMP uses provenance for training-time credit assignment, whereas PANINI-DR is training-free and uses its evidence and review records for measurement and control.An admitted binding records the proposal, attached evidence, and review outcome.Source: §E.4 · p. 37

Related papers

Loading 2610.15208v1…