PaperlayerView exact revision on arXiv

Source-linked AI summary

Revision-Aware Independent Agent Graphs for Dynamic Reasoning

Yan Luo, Selim-Antoine Lali, Jeremy Moebel, Iliass Khoutaibi, Ahmadou Aidara, Mengyu Wang

arXiv:2610.01249v1cs.AIcs.CL

TL;DR

Fixed-task evaluations do not test whether agents track task bindings revised by event histories. This paper defines dynamic task routing and introduces RIAG; across six datasets, homogeneous RIAG achieves 54.24% joint accuracy at 0.62 calls per query, versus Debate’s 32.22% at 18.00 calls per query.

  • Problem

    Conventional evaluations supply a fixed task, leaving untested whether agents select the valid document after delayed, scheduled, superseded, expired, or retracted instructions.

    Source: §1 · p. 2

  • Method

    RIAG deterministically resolves the task at query time, reuses solutions only for exact document-version and configuration matches, and conditionally audits and repairs independent proposals.

    Source: §1 · p. 3

  • Results

    54.24% joint accuracy at 0.62 calls/q for homogeneous RIAG compares with Debate’s 32.22% joint accuracy at 18.00 calls/q across six datasets.

    Source: §5.1 · p. 7

  • Takeaways & Limitations

    The results indicate that reliable temporal binding and selective reuse can matter more than repeated deliberation or model diversity alone.

    Source: §1 · p. 3

    The collection uses one seed, checkpoints within an episode are correlated, and isolating mechanisms requires matched controls, multiple seeds, and randomized post-trigger actions.

    Source: §D.4 · p. 14

Abstract

from arXiv · show

Conventional reasoning protocols present a fixed, preselected task, so they cannot test whether an agent propagates relevant updates, preserves unaffected work, or reconstructs a historical task binding. We therefore study \emph{dynamic task routing}, in which an event stream revises task bindings and a system must select the document version valid at each query time before solving it. To study this problem, we repurpose six widely used benchmarks: MMLU, MMLU-Pro, MedMCQA, MATH, GPQA, and HumanEval into 31{,}119 dynamic episodes comprising 373{,}428 temporally categorized queries. This setting exposes a central trade-off: recomputing after every event wastes work, whereas unguarded reuse returns stale conclusions. We introduce the Revision-Aware Independent Agent Graph (RIAG), a bounded multi-agent policy that separates deterministic temporal resolution from task reasoning. RIAG caches solutions by immutable document identity, starts each fresh task with two unexposed attempts, and conditionally invokes audit and repair, using at most four calls per document version. On this collection, homogeneous RIAG achieves 54.24\% joint routing-and-answer accuracy at 0.62 calls/query, compared with 32.22\% at 18.00 calls/query for the strongest comparison method; heterogeneous RIAG reaches 49.78\% at 0.63 calls/query.

1 INTRODUCTION

The introduction frames dynamic task routing as reasoning over task bindings revised by an event stream, rather than solving a fixed, preselected task. It introduces RIAG and evaluates it across six repurposed benchmarks.

1 INTRODUCTION

  • The motivating problem is to decide what to recompute or reuse as events revise the active task, and whose agreement counts as evidence.A wrong binding means reasoning about a superseded task, while recomputation after every event wastes calls and latency on unchanged inputs.Source: §1 · p. 1
  • Dynamic task routing evaluates whether agents select the document valid at each query time from a revisable event history, then solve it.The setting repurposes MMLU, MMLU-Pro, MedMCQA, MATH, GPQA, and HumanEval while preserving their questions and evaluators.Source: §1 · pp. 2–3
  • RIAG resolves the valid task deterministically, reuses solutions only for exact document-version and configuration matches, and counts agreement only before agents communicate.A cache miss starts with independent solver and checker attempts; targeted audit and repair are conditional, with at most four calls for a fresh task.Source: §1 · pp. 2–3
  • Across 31,119 episodes and 373,428 scheduled queries per method, RIAG achieved 54.24% joint accuracy versus Debate’s 32.22%, using 0.62 rather than 18.00 model calls per query.Source: §1 · p. 3
  • The study spans homogeneous and specialized model pools and eight comparison families, and examines temporal binding, reuse, inference settings, and response diversity.The authors report that greater response diversity does not necessarily improve final accuracy.Source: §1 · p. 3

2 RELATED WORK

The related work positions RIAG alongside graph orchestration, multi-agent deliberation, and agent memory systems. Its graph records exposed outputs while combining task resolution with exact-key solution reuse.

Graph-structured agent orchestration

  • Unlike workflow-search graphs, RIAG’s vertices are model invocations and its edges record which outputs later calls saw, producing an auditable execution trace.It combines that trace with temporal document resolution, version-keyed caching, and a fixed four-call budget.Source: §2 · p. 3

Multi-agent deliberation and specialization

  • Where deliberation systems organize communication after task selection, RIAG first resolves the valid task and preserves two unexposed proposals as independent evidence.Audit and repair follow only disagreement or structural failure; post-exposure agreement is treated as dependent evidence.Source: §2 · p. 3

Agent memory, reflection, and revision

  • RIAG reuses a result only when the task document and inference configuration match exactly, resolving the active document before consulting its cache.A route returning to an earlier document can reuse its answer, but an answer is never carried over to a different document.Source: §2 · p. 4

3 DYNAMIC TASK ROUTING PROTOCOL

The protocol turns fixed benchmark questions into episodes where event histories determine which fixed-content task document applies at each checkpoint. It tests route preservation, propagation, historical revision, and solution reuse.

3 DYNAMIC TASK ROUTING PROTOCOL

  • Delayed reports and future-effective events should not displace the current route; expiration restores an underlying route, while retraction can revise a historical binding.The example’s delayed report is effective before a later alert, and a future-effective instruction changes the route only when its effective time arrives.Source: §3 · p. 4
  • At each checkpoint, the system reconstructs the valid route from evidence available at query time before solving the selected document.Task content stays fixed as an experimental control; a changed payload would receive a new immutable version and answer key.Source: §3 · p. 4
  • The protocol separately tests preservation, propagation, and historical revision by checking delayed or early events, route restoration after expiration, and retractions.A historical query can resolve differently after a previously trusted event is retracted.Source: §3 · p. 4
  • The collection repurposes six benchmarks into 31,119 episodes, each with three task documents, 12 events, and 12 checkpoint queries.The checkpoints comprise one baseline, one changed, four preservation, three propagation, and three historical queries.Source: §3 · p. 4
  • Selection accuracy, answer accuracy, and joint accuracy distinguish correct routing from solving the gold task and achieving both on the same query.Answer-only evaluation can credit the correct answer string even when the system follows the wrong task.Source: §3 · p. 5

4 REVISION-AWARE INDEPENDENT AGENT GRAPH

RIAG separates temporal task selection from task reasoning, then reuses solutions only for matching document versions and inference configurations. On a cache miss, its bounded graph starts with unexposed proposals and conditionally invokes audit and repair.

4.1 PROBLEM SETUP

  • Queries target fixed task documents over an event history; current queries use world time equal to knowledge time, while historical queries use an earlier world time.The admissible history excludes future events, gold answers, and hidden tests.Source: §4.1 · p. 5

4.2 OVERVIEW

  • A fresh cache key opens a graph with two unexposed proposals, followed by an auditor on disagreement or invalid output and at most one repair; cache hits cost no calls.A fresh key costs two to four calls, and every call records its exposure ancestry.Source: §4.2 · p. 5, §4.5 · pp. 6–7

4.3 WHICH TASK? SELECTING THE DOCUMENT IN FORCE

  • At each query, a deterministic resolver selects the document version valid at the requested world time from the visible event history, without a model call.Known, effective retractions can revise later answers even about earlier world times.Source: §4.3 · p. 6

4.4 RECOMPUTE OR REUSE? VERSION-KEYED MEMOIZATION

  • RIAG reuses a solution only when the document version and inference configuration match; a changed version or configuration forces recomputation.Leaving and later returning to the same version reuses its solution; the episode-local cache is cleared at episode start.Source: §4.4 · p. 6

4.5 WHOSE AGREEMENT COUNTS? A BOUNDED, EXPOSURE-AWARE GRAPH

  • Solver and checker commit before seeing each other's output, while auditor and repairer review exposed records, so later judgments are not additional independent votes.Format checks do not test semantic correctness, and a valid repair is unresolved because no later call verifies it.Source: §4.5 · pp. 6–7

4.6 INSTANTIATIONS

  • The evaluated role-to-model mapping is part of the inference configuration, so changing any role's model invalidates reuse.Homogeneous RIAG uses Qwen2.5-7B-Instruct for every role; the heterogeneous mapping assigns different models to solver, checker, auditor, and repairer.Source: §4.6 · p. 7

5 EXPERIMENTS

Across six dynamic task-routing datasets, RIAG achieves higher joint accuracy than comparison methods while using fewer calls. The experiments also examine temporal robustness, resolver transfer, reuse, and response diversity.

5.1 OVERALL RESULTS

  • Overall Joint is a query-weighted aggregate of five categories; selection accuracy requires the gold document, answer accuracy the correct gold-task answer, and joint accuracy both on one query.Preservation queries receive the largest category weight.Source: §5.1 · p. 7

5.2 TEMPORAL BEHAVIOR, CONTROLS, AND ROBUSTNESS

  • 30.07- and 30.63-point homogeneous, and 19.95- and 27.73-point heterogeneous joint-accuracy gains occur when the deterministic resolver controls GoA on MATH.Selection rises from 58.30–61.12% for vanilla GoA to 99.35–99.87% across the four pooling and model-pool combinations.Source: Figure 3 · p. 8, §5.2 · p. 8

5.1 OVERALL RESULTS

  • 54.24% joint accuracy, 100.00% selection accuracy, and 0.62 calls/q for homogeneous RIAG, versus Debate’s 32.22% joint accuracy at 18.00 calls/q.Heterogeneous RIAG reaches 49.78% joint accuracy, 98.81% selection accuracy, and 0.63 calls/q.Source: §5.1 · p. 7, Table 1 · p. 8

5.2 TEMPORAL BEHAVIOR, CONTROLS, AND ROBUSTNESS

  • Homogeneous RIAG stays within 53.89–54.50% joint accuracy across query categories, while conventional methods lose 24–29 points from baseline to historical queries.Across six datasets, homogeneous RIAG ranges from 40.10–66.83% joint accuracy, versus 23.02–38.66% for the best comparison method on each dataset.Source: §5.2 · p. 8

5.3 REUSE, CONDITIONAL EXPANSION, AND RESPONSE DIVERSITY

  • 16.4% to 32.0%: useful diversity rises, while initial answer coverage falls from 73.4% to 68.8% and final unique-task accuracy from 67.6% to 57.5%.Coverage also declines from the closest to farthest distance quartile; audit rescue and damage rates are descriptive because difficult cases trigger auditing.Source: Figure 4 · p. 9, §5.3 · p. 9

5.2 TEMPORAL BEHAVIOR, CONTROLS, AND ROBUSTNESS

  • RIAG’s sensitivity sweep yields 62.05–64.50% joint accuracy for the homogeneous pool and 49.70–57.02% for the heterogeneous pool.The 1,024-token homogeneous limit is associated with a 2.08-point decrease and 115 format failures; heterogeneous temperature 1.0 with a 5.18-point decrease and no format failures.Source: §5.2 · p. 9

5.3 REUSE, CONDITIONAL EXPANSION, AND RESPONSE DIVERSITY

  • On reference MATH runs, 76.9% of queries are cache hits; fresh-task two-call agreement occurs in 66.2% of homogeneous and 41.5% of heterogeneous cases.Audit rates are 30.2% and 51.3%, and repair rates are 3.5% and 7.2%, respectively. Reuse follows event semantics: all preservation and historical queries reuse solutions, as do 74.3% of propagation queries; fresh-task agreement routes are accurate in 82.1% and 85.4% of cases.Source: §5.3 · p. 9

6 CONCLUSION

The paper concludes that RIAG separates temporal task resolution from task reasoning, enabling reuse across event revisions while counting agreement only before communication. It also discloses language-model use in writing and dataset curation.

6 CONCLUSION

  • RIAG separates temporal task resolution from task reasoning, reuses solutions across event revisions, and counts agreement only when it precedes communication.The conclusion reports higher joint accuracy at fewer than one model call per query on six datasets and resolver-controlled GoA gains with a different reasoning-graph family.Source: §6 · p. 10

A USE OF LARGE LANGUAGE MODELS

  • GPT-6 Astra assisted with grammar, text refinement, and dataset curation, but did not change source questions, options, answer keys, code entry points, or hidden tests.Episodes were generated by a deterministic builder and checked by three human reviewers and an independent validator.Source: A USE OF LARGE LANGUAGE MODELS · p. 12

B REPRODUCIBILITY STATEMENT

The paper provides code and a deterministic benchmark collection, describes its evaluation protocol, and states ethical scope boundaries.

B REPRODUCIBILITY STATEMENT

  • The codebase and collection are supplementary materials; a seeded deterministic builder creates the collection from six public benchmarks.The paper points to appendices for curation protocol, statistics, model checkpoints, decoding settings, and context lengths.Source: B REPRODUCIBILITY STATEMENT · p. 12

C ETHICS STATEMENT

  • The data use public benchmarks and synthetic events that change only which task is active; RIAG inherits model biases and errors, so medical answers are not for clinical decisions.Source: C ETHICS STATEMENT · p. 12

D.1 EVALUATION PROTOCOL

  • Each method is evaluated once with seed 0 across 31,119 episodes and 373,428 scheduled queries spanning six benchmarks.Methods include eight multi-agent comparisons; all use temperature 0.7, top-p 1.0, and a 2,048-token output cap, while baseline context is 16,384 tokens and GoA/RIAG context is 32,768.Source: §D.1 · p. 13
  • Selection accuracy measures gold-document routing, answer accuracy measures correctness for the gold task, joint accuracy requires both, and calls/q counts recorded calls per scheduled query with cache hits at zero.Source: §D.1 · p. 13

D.2 MAIN RESULTS

RIAG outperforms the strongest comparison method on joint accuracy while using far fewer model calls. The reported results also show strong selection accuracy, but evaluate RIAG as a complete system rather than isolating its graph policy.

D.2 MAIN RESULTS

  • 54.24% joint accuracy at 0.62 calls/query for homogeneous RIAG exceeds Debate's 32.22% at 18.00 calls/query; heterogeneous RIAG reaches 49.78% joint accuracy.RIAG selection accuracy is 100.00% for the homogeneous pool and 98.81% for the heterogeneous pool.Source: §D.2 · p. 13, Table 2 · p. 13

D.3 PARALLEL INDEPENDENT CALLS AND RECORDED COST

  • The full-collection runs used parallel independent calls on 16 workers, but provide no serial-versus-parallel speed-up or accuracy-equivalence comparison.The homogeneous totals include recovery calls from a repaired overlay; the heterogeneous run retains parse-failed queries counted as incorrect.Source: §D.3 · p. 14, Table 3 · p. 14

D.4 INTERPRETATION AND LIMITATIONS

  • These experiments assess RIAG as a complete system, not the isolated effect of its graph policy.The systems differ in resolver, reuse, model configuration, answer path, decoding constraints, and context length; the collection uses one seed and MATH uses exact-string scoring.Source: §D.4 · p. 14

D.5 DATASET AND QUERY-CATEGORY BREAKDOWNS

  • On MATH, homogeneous and heterogeneous RIAG reach 63.20% and 54.50% joint accuracy, while no conventional baseline exceeds 29.60%.RIAG selection accuracy is 100.00% and 98.81%, respectively, but better routing alone does not solve the task.Source: §D.5 · p. 14

D.6 MATH ANALYSIS

The MATH analysis examines resolver-controlled Graph-of-Agents and RIAG sensitivity to inference settings. Resolver-controlled GoA improves task binding and joint accuracy in the tested configurations, while the one-factor sweep shows that output budget and sampling can affect accuracy.

D.6 MATH ANALYSIS

  • The MATH sweep crosses nine one-factor settings with homogeneous and heterogeneous pools across 18 configurations, 500 episodes each.The reference uses temperature 0.7, top-p 1.0, a 2,048-token output cap, and seed 0; seed changes measure inference sensitivity, not independent training replications.Source: §D.6 · p. 14

D.6.1 RESOLVER-CONTROLLED GRAPH-OF-AGENTS

  • Resolver-controlled GoA raises selection from 58.30–61.12% to 99.35–99.87% and improves joint accuracy across all four tested settings.Joint-accuracy gains are 30.07 and 30.63 points for homogeneous max and mean pooling, 19.95 for heterogeneous max, and 27.73 for heterogeneous mean.Source: §D.6.1 · pp. 15–16
  • The resolver-controlled GoA comparison is not a one-factor ablation: role cards and greedy decoding differ from vanilla GoA, and each cell has one run.The observed gap therefore belongs to the complete resolver-controlled package.Source: §D.6.1 · p. 16

D.6.2 SENSITIVITY TO DECODING AND OUTPUT BUDGET

  • A 1,024-token homogeneous cap changes accuracy by −2.08 points (95% CI [−3.85, −0.33]); heterogeneous temperature 1.0 changes it by −5.18 points ([−7.83, −2.58]).Most paired intervals include zero; the temperature effect occurs without format failures, indicating sampling sensitivity rather than truncation.Source: §D.6.2 · p. 16
  • The homogeneous pool is stable over most tested settings, but neither pool is invariant to output budget or sampling, and the one-factor design does not identify interactions.Greedy-decoding point estimates are positive in both pools, but both intervals include zero.Source: §D.6.2 · p. 16

D.6.3 ACCURACY AND RECORDED COMPUTE

The MATH analysis compares accuracy and recorded compute across RIAG pools and inference settings, then examines temporal reuse and conditional graph expansion. Cache hits account for most queries, while fresh-task routes vary in how often they proceed to audit or repair.

D.6.3 ACCURACY AND RECORDED COMPUTE

  • The reference homogeneous pool reaches 64.13% accuracy at 0.548 calls and 285.9 output tokens/query; the heterogeneous pool reaches 54.88% at 0.614 calls and 318.5 tokens/query. Here, heterogeneity costs more and performs worse.Source: §D.6.3 · p. 16
  • Greedy homogeneous decoding uses the fewest calls and is slightly more accurate than its reference, while a 1,024-token homogeneous cap reduces tokens and accuracy but increases calls/query.For the heterogeneous pool, temperature 1.0 produces the lowest accuracy and highest call and token counts; a larger output cap adds tokens without a matching gain.Source: §D.6.3 · p. 16, Figure 5 · p. 17

D.6.4 TEMPORAL REUSE AND CONDITIONAL GRAPH EXPANSION

  • In both reference pools, 4,614 of 6,000 queries (76.9%) are cache hits; all preservation and historical queries reuse a solution.Baseline and changed queries always trigger fresh reasoning, while 74.3% of propagation queries reuse a solution.Source: §D.6.4 · p. 17
  • Among 1,386 fresh tasks, homogeneous RIAG stops after two agreeing calls on 66.2%, invokes the auditor on 30.2%, and reaches repair on 3.5%.Heterogeneous rates are 41.5%, 51.3%, and 7.2%; these conditional accuracies are not intervention effects because disagreement or failed validation selects tasks into deeper routes.Source: §D.6.4 · p. 17

D.6.5 REASONING-SPACE DIVERSITY AND CONDITIONAL INTERVENTION

This section examines response diversity and evaluation in dynamic task routing, emphasizing that diversity alone does not guarantee correctness and identical answers can conceal selection of the wrong document. It treats embedding distances as descriptive evidence and motivates reporting document-selection accuracy, answer accuracy, and their conjunction.

D.6.5 REASONING-SPACE DIVERSITY AND CONDITIONAL INTERVENTION

  • A four-class role probe reaches 34.3% balanced accuracy for homogeneous RIAG and 65.7% for heterogeneous RIAG, where role and checkpoint are confounded.Initial solver/checker responses predict the pool at 67.5% balanced accuracy, indicating model assignment changes response geometry before communication.Source: §D.6.5 · p. 18
  • Mean solver–checker cosine distance is 0.190 for heterogeneous RIAG and 0.111 for homogeneous RIAG, with a paired difference of 0.080 (95% CI [0.075, 0.084]).Heterogeneous distances were larger on 83.8% of tasks, with a standardized effect of 0.85; this establishes response diversity, not better reasoning.Source: §D.6.5 · p. 18, Figure 7 · p. 19
  • Useful diversity rises from 16.4% to 32.0% and wrong agreement falls from 9.5% to 4.6%, yet final normalized accuracy is 67.6% for homogeneous and 57.5% for heterogeneous RIAG.Initial answer coverage falls from 73.4% to 68.8%; these unique-task normalized-string results do not replace scheduled-query exact match.Source: §D.6.5 · p. 18
  • From the closest to farthest distance quartile, initial answer coverage drops from 80.7% to 62.8% for homogeneous RIAG and from 79.8% to 55.9% for heterogeneous RIAG.In both pools, wrong agreements are closer than disagreeing pairs; heterogeneous both-wrong, disagreeing pairs are most separated, so distance signals disagreement but not productive diversity.Source: §D.6.5 · pp. 18–19, Figure 8 · p. 19
  • Auditing yields conditional rescue rates of 2.5% for homogeneous and 0.6% for heterogeneous tasks, damage rates of 19.2% and 19.4%, and final normalized accuracies of 32.8% and 36.5%.Disagreement or failed validation triggers auditing, so these descriptive rates are not causal effects; causal estimation requires randomizing the post-trigger action.Source: §D.6.5 · pp. 19–20

E.2 SOURCE COVERAGE AND EPISODE CONSTRUCTION

  • The collection converts six public benchmarks into 31,119 episodes, but MMLU and MMLU-Pro account for 83.8% of episodes, so aggregate scores require per-dataset results.The scope is every row of six named local exports, not every upstream split; each episode contains 12 checkpoint queries.Source: §E.2 · p. 22, Table 9 · p. 22

E.3 MODEL INPUTS, OUTPUT CONTRACTS, AND SCORING

  • Each episode uses three fixed-content task documents, and RIAG resolves the selected document from events received so far before solving it, without gold feedback.Historical, delayed, future-effective, superseded, expired, and retracted assertions follow explicit temporal rules; task content and answer keys remain fixed while events change bindings.Source: §E.3 · pp. 23, 25

E.4 DIAGNOSTIC COMPOSITION AND COMPARISON

  • The 373,428 queries comprise 31,119 baseline, 31,119 changed, 124,476 preservation, 93,357 historical, and 93,357 propagation queries.Preservation tests resistance to irrelevant or not-yet-effective events, propagation tests relevant changes, and historical queries test reconstruction from evidence available at arrival.Source: §E.4 · p. 25
  • Dynamic task routing creates 12× as many scored checkpoints and 26× as many public input records as one isolated prompt and answer per source item.Visible text fields grow from 2.70 million whitespace-separated tokens in original prompts to 30.05 million in dynamic streams, an 11.1× increase; this is a workload descriptor, not a tokenizer count or an accuracy or sample-efficiency gain.Source: §E.4 · p. 25
  • 19.1% of exact-answer episodes have the same baseline and changed answer string despite routing to different documents, showing why answer-only scoring is insufficient.The collision rate is 5,904 of 30,955 exact-answer episodes.Source: §E.4 · p. 25
  • Collision rates are 25.3% for MMLU, 10.0% for MMLU-Pro, 26.2% for MedMCQA, 1.0% for MATH, and 25.8% for GPQA; HumanEval is excluded.HumanEval is excluded because its response is executable code evaluated with hidden tests.Source: §E.4 · p. 25

E.5 SEMANTIC COVERAGE AND EPISODE CONSTRUCTION

The source tasks overlap semantically, and the episode constructor shows no evidence of systematically selecting partners with atypical semantic separation. Outcome maps are descriptive diagnostics; the method details specify bounded calls, exact cache identity, and what its guarantees do not cover.

E.5 SEMANTIC COVERAGE AND EPISODE CONSTRUCTION

  • Dataset-label silhouette was 0.033, with MMLU and MMLU-Pro nearest neighbors substantially intermixed; UMAP island boundaries are unstable across seeds.Their cross-neighbor fractions were 35% for MMLU and 37% for MMLU-Pro; alternate-seed 15-neighbor Jaccard agreement was 0.332.Source: §E.5 · pp. 26–27, Figure 11 · p. 27
  • Observed episode distances nearly matched within-dataset random-partner controls, with paired Wilcoxon p=0.543 and mean differences no larger than about 0.01.This provides no evidence of systematic partner-distance selection, but does not establish that semantic distance is irrelevant to difficulty.Source: Figure 12 · p. 27, §E.5 · pp. 27–28
  • The 31,119 source tasks contain 1,365 rows in 682 exact-text duplicate groups, including 420 cross-dataset groups with 841 rows.The passage recommends grouping duplicates when constructing future training and test splits; duplicates do not alter this evaluation.Source: §E.5 · p. 27

E.6 OUTCOME LANDSCAPES IN SEMANTIC SPACE

  • Heterogeneous minus homogeneous RIAG joint accuracy was −2.69 points in the complete-run snapshot, while homogeneous GoA mean exceeded max pooling by 0.76 points overall.These are descriptive artifacts preceding the repaired homogeneous RIAG aggregate, not replacements for the main result tables.Source: Figure 13 · p. 28, §E.6 · p. 28
  • Outcome maps use 50-neighbor averages in the original embedding space, displayed on UMAP coordinates, and do not establish causal sources of performance differences.The maps cover complete-run homogeneous GoA max and mean, and homogeneous and heterogeneous RIAG; heterogeneous GoA runs were not included.Source: §E.6 · p. 28, Figure 13 · p. 28
  • Across four complete runs, selection accuracy stayed nearly flat across episode-distance quartiles, and joint-accuracy method ordering persisted across quartiles.These are descriptive aggregates over different dataset mixtures, not an intervention; UMAP axes and long-range distances are not quantitative semantic measures.Source: Figure 14 · p. 29, §E.6 · p. 29

F.1 COST AND GUARANTEES

  • Episode call cost is 2n2 + 3n3 + 4n4 ≤ 4U, while cache hits cost no calls; the stated invariants do not guarantee answer correctness.Prefix isolation, fresh resolution, exact-key version binding, exposure accounting, and budget integrity are enforced; unversioned in-place edits are excluded and require a new version and key.Source: §F.1 · p. 29

F METHOD DETAILS

  • Invocation records retain role, cache key, document event, model and prompt revisions, sampling seed, parsed answer, failures, status, token use, and peer-text hashes.Format checks validate answer structure rather than factual correctness; a well-formed wrong answer or buggy program can pass.Source: §F.1 · p. 30

F.2 A SHARED-EPISODE RESPONSE EXAMPLE

The shared GPQA episode illustrates why scoring must require both the routed document and its answer: stale routing can earn a coincidental answer-symbol match. RIAG follows the superseding event to the correct document, unlike the shown GoA variants.

F.2 A SHARED-EPISODE RESPONSE EXAMPLE

  • At Q2, the delivery route follows the time-8 supersession from C to B, while the archive request remains bound to C.Table 11 presents the full 12-query stream and six recorded responses for this example.Source: §F.2 · p. 30, Table 11 · p. 31
  • Both RIAG variants followed the time-8 superseding event to document B and returned its gold answer D; all four GoA variants instead selected stale document C.The complete response is scored as a (document, answer) pair, so matching B’s answer symbol while naming C is incorrect.Source: §F.2 · pp. 30–31, Table 11 · p. 31
  • Three GoA variants returned answer D while labeling the stale document C, although C’s own gold answer is C.This is a coincidental option-symbol match across different tasks, not a correct response.Source: §F.2 · p. 31
Loading 2610.01249v1…