Controlled Acquisition and Abstention in Three-Channel Score Conflicts

1 INTRODUCTION

• The benchmark asks when a third source is worth acquiring, when conflict should trigger abstention, and whether request value can be predicted from the observed pair.Acquisition, incorrect answers, delays, and abstention can carry different application-dependent costs.Source: §1 · pp. 1–2
• A policy observes two signed scores, may request the third, and answers when two strong scores agree; otherwise it abstains.Training and evaluation use separately generated samples, with partly held-out corruption mechanisms.Source: §1 · p. 2

1.1 CONTRIBUTIONS

• The three-score protocol separates acquisition from the terminal answer or abstention decision, and matched controls test when request savings leave final actions unchanged.The common-rule, matched-budget test compares train-only value selection with pair uncertainty, matched random, and a non-deployable oracle.Source: §1.1 · p. 2

2 RELATED WORK

• Representation-fusion framing is incomplete for conflicts because systems must assess example-specific source reliability; modality-interference and credibility-aware work study related issues [Tenali et al., 2026].CrossCheck-Bench and AV-ConfuseBench use controlled perturbations for image–text contradiction and audiovisual confusion [Tian et al., 2026].Source: §2 · p. 3
• EmoMM steers attention across supplied audio, video, and text under emotion conflict and missingness, whereas this study prices requests and permits abstention [Sun et al., 2026].Its evidence is generated scalar scores with bounded frozen-feature media diagnostics, unlike EmoMM's media-model evaluation.Source: §2 · p. 3
• C3PO examines counterfactual conflict across four modalities; this study instead uses a three-score generator and partially held-out split to make action semantics explicit [Mukherjee et al., 2026].The narrower setup trades media realism for tests of observation-only decisions and error sources.Source: §2 · p. 3
• Selective prediction and active feature acquisition cover rejection or purchasing missing information, while this setting fixes two observed scores, one unavailable score, and separate request and abstention decisions [Geifman & El-Yaniv, 2019] [Li & Oliva, 2021] [Li & Oliva, 2025].Related work also includes temporal acquisition controls and expected downstream loss reduction [Kossen et al., 2023] [Valancius et al., 2024], and iterative visual question asking [Liu et al., 2026].Source: §2 · p. 3
• The synthetic targeted decision accuracy rewards abstention only on a designated ambiguity mechanism, so comparisons with standard selective risk require the target convention to be stated.The paper says its evidence does not establish a general acquisition method.Source: §2 · p. 3

3 PROBLEM FORMULATION

• Each example has a binary latent event and three audio, video, and text scores; corruption mechanisms modify one or more scores, and the partially held-out split introduces mixed, burst, and ambiguity cases.The seven mechanism labels correspond to six distinct score-level transformations because invert and swap use the same operation.Source: §3 · p. 3
• Ambiguity receives an abstention target, but this mechanism-specific convention is not a general definition of non-identifiability.Mixed and ambiguity are both observation-non-identifiable with balanced labels, yet the reward charges abstention on mixed and rewards it on ambiguity.Source: §3 · p. 4
• The terminal action is a binary label or abstention; correctness is measured against the latent label on non-ambiguous cases and against abstention on ambiguity cases.The policy's reward also depends on whether it requests the third score.Source: §3 · p. 4
• Utility gives correct actions +1 and incorrect actions −1 before costs, with a 0.10 request cost and a 0.20 cost for non-ambiguous abstention.Targeted decision accuracy uses the mechanism-conditioned target.Source: §3 · p. 4

4.1 GENERATOR AND SPLITS

• Each scalar run uses 6,000 training examples and 4,000 evaluation examples; the partially held-out split adds mixed, burst, and ambiguity cases, with only clean cases of those types appearing in training.Training and seen splits sample clean, invert, swap, and dropout; invert and swap share a transformation and thus receive twice the sampling and mechanism-macro weight.Source: §4.1 · p. 4

4.2 FIXED AND LEARNED BASELINES

• Full-information majority, median, weighted, MLP, and gated policies receive all three scores initially, while learned acquisition is trained to predict whether requesting improves utility.The learned scorer sees only the initial scores and magnitudes at fitting and inference; its threshold is chosen on a separate train-only calibration split.Source: §4.2 · p. 4

4.3 THRESHOLD QUERY-AND-ABSTAIN POLICY

• THRESHOLD requests when the initial signs disagree or either score has magnitude below 0.35, then answers only if two scores are strong and share a sign.Its threshold was fixed before held-out evaluation, and the rule does not inspect mechanism labels or ground truth.Source: §4.3 · p. 5

4.4 MATCHED PROTOCOL CONTROLS

• The threshold rule and always-query-selective have identical final decisions; across five seeds their 0.03656 utility gap equals the cost of avoided requests.Query-majority has the same request rate but lower utility.Source: §4.4 · p. 5

4.5 MATCHED-BUDGET SELECTOR TEST

• The matched-budget selector uses a training-only target for the reward change from adding the third score and ranks evaluation pairs using only the initial pair.It is compared with pair uncertainty, seed-specific random ranking, and an oracle at 10%, 25%, 50%, and 63.7% budgets; the oracle uses unavailable evaluation information.Source: §4.5 · p. 5

5 EXPERIMENTAL SETUP

• The temporal evaluation uses seven equally likely mechanisms, so its expected ambiguity fraction differs from the scalar evaluation and their accuracy levels are not directly comparable.Neither temporal model abstained; their three-class heads receive no ambiguity examples during training, and no mechanism labels or ambiguity flags are inputs.Source: §5 · p. 5
• At T = 64, equal-information temporal controls use the same full traces and test central and hard settings with shared, independent-channel, and per-example position perturbations.The study tests whether the initial separation requires the fixed waveform or temporal order.Source: §5 · p. 5
• The scalar references receive all three scores without request costs, whereas acquisition policies start with two; the studies use different analysis units, and actor allocations overlap.The AMI evaluation covers 244 donor-matched dialogue-act cases across two teams.Source: §5 · p. 6
• The partially held-out split adds mixed corruption, high-confidence bursts, and symmetric ambiguity while retaining clean cases; utility is not tuned on held-out mechanisms.The temporal factorial has nine sequence-length/noise cells and 207 setting-seed runs; temporal models receive full traces while scalar controls receive time means.Source: §5 · pp. 5–6

6 RESULTS

• On the partially held-out split, threshold reaches 0.789 ± 0.006 targeted decision accuracy and 0.481 ± 0.014 utility, versus majority’s 0.626 ± 0.008 and 0.252 ± 0.016.Majority receives all three scores, while threshold starts with two; Table 1 also reports non-ambiguous binary accuracy of 0.721 ± 0.008 for threshold versus 0.834 ± 0.007 for majority.Source: §6 · p. 6, Table 1 · p. 7, Figure 2 · p. 7

6.1 MECHANISM-LEVEL BEHAVIOR

• On novel ambiguity, the active policy abstains at 0.992; on burst cases it retains 0.977 accuracy and queries on 0.674 of examples, while mixed-case utility is −0.839.The threshold rule covers 0.590 of cases and has 0.918 ± 0.005 accuracy when answering; if all abstentions are errors, its utility is −0.064 ± 0.019 versus majority’s 0.252 ± 0.016.Source: §6.1 · pp. 6, 8

6.2 TEMPORAL SIMULATION RESULTS

• At T = 64 and the central setting, full-trace template and ridge controls match the networks at 0.786 targeted decision accuracy and 0.572 utility; position perturbations separate them.With independent channel permutations, fusion retains 0.768 ± 0.009 targeted accuracy while ridge and template reach 0.428 ± 0.015 and 0.428 ± 0.008, respectively; this remains simulator-specific.Source: §6.2 · p. 8, Table 2 · p. 8

6.3 WHOLE-POLICY LEARNED ACQUISITION DIAGNOSTIC

• On the partially held-out split, learned acquisition queries less often than threshold but reaches 0.746 ± 0.010 targeted decision accuracy and 0.417 ± 0.019 utility versus 0.789 ± 0.006 and 0.481 ± 0.014.The scorer’s no-request branch changes the terminal answer rule, so this whole-policy comparison cannot isolate the value of its query schedule.Source: §6.3 · p. 9

6.4 COST SENSITIVITY AND INFORMATION BUDGET / 6.5 MATCHED CONTROLS AND THRESHOLD SENSITIVITY

• At campaign costs (0.40, 0.80), active utility is 0.192 ± 0.018, below majority’s 0.252 ± 0.016; at the matched control, threshold and always-query-selective make identical final decisions.Their utility gap is 0.03656, equal to avoided-query cost 0.10(1 − 0.6344); excluding shared-clean cases, active targeted accuracy is 0.720 ± 0.006 and utility is 0.313 ± 0.013.Source: §6.4 · p. 9, §6.5 · p. 9

6.6 MATCHED-BUDGET SELECTOR VALUE

• At 10% and 25% budgets, train-only value selection exceeds no-query utility by 0.0138 and 0.0041, but learned utility is lower than pair uncertainty at 50% and 63.7%.At those higher budgets, learned requests have mean reward changes of −0.0591 and −0.1343 before query cost, versus +0.1756 and +0.0235 for pair uncertainty.Source: §6.6 · p. 9

6.7 ORACLE VALUE VERSUS PRE-QUERY CUE

• Oracle headroom does not establish feasible pre-query value; on CMU-MOSEI descriptors, router error-reduction difference is 0.000499, below the prespecified 0.005 margin.The Perception Test pool falls below the source-grounding and unseen-component thresholds for a conflict benchmark or held-out evaluation.Source: §6.7 · pp. 9–10

6.8 NATURAL-MEDIA DIAGNOSTICS

• On 1,244 actor-held-out CREMA-D clips, eight-frame fusion minus audio accuracy is +5.87% for one encoder pair and −2.49% for another; both actor intervals contain zero.Matched-request routing gains are small and uncertain; neither study establishes a reliable selective-acquisition gain.Source: §6.8 · p. 10

• The threshold rule saves requests under its synthetic reward.Source: §7 · p. 10
• Learned selection helps only at low budgets and loses to pair uncertainty in utility.Source: §7 · p. 10
• Natural-media acquisition gains, latency, privacy, and perceptual benefit remain unconfirmed; human evaluation is future work.Source: §7 · p. 10

• AI tools helped polish the writing.Source: AI USE STATEMENT · p. 11
• AI tools also helped find related work.Source: AI USE STATEMENT · p. 11
• The paper describes details of the AI-tool use.Source: AI USE STATEMENT · p. 11

REPRODUCIBILITY STATEMENT

• The project repository provides the ConflictBench implementation and setup instructions, while project-owned code is released under the MIT License.Dataset and dependency terms remain with their respective providers.Source: REPRODUCIBILITY STATEMENT · p. 11

A EXTENDED TRANSFER AND SOURCE DIAGNOSTICS

• The appendix distinguishes screen designs and study units, including descriptor routing, a training-only source study, and actor- or team-held-out media evaluations.The reported counts refer to different objects, including utterances, videos, actors, and meeting cases, and are not interchangeable with synthetic seeds or temporal settings.Source: A EXTENDED TRANSFER AND SOURCE… · p. 14

A.1 CMU-MOSEI DESCRIPTOR TRANSFER

• CMU-MOSEI transfer uses projected descriptors and official folds, without target-domain threshold tuning or a synthetic ambiguity target or corruption.The fixed threshold is τ = 0.35; the benchmark results are not a comparison with current multimodal encoders.Source: §A.1 · p. 14
• 0.624 test accuracy for the threshold rule trails majority at 0.653 and MLP at 0.773 ± 0.002 on CMU-MOSEI descriptors.Table 3 reports official-fold descriptor accuracy; the calibrated rule reaches 65.17% accuracy and 0.2659 utility, while same-request majority reaches 65.29% and 0.2689.Source: §A.1 · pp. 14–15, Table 3 · p. 15

A.2 PROJECTED-DESCRIPTOR ACTION SCREEN

• The projected-descriptor action screen finds no supported acquisition advantage: intervals cross zero, and the value router misses its preregistered margin over matched active baselines.At the 0.50 target budget, the second screen's router margins are 0.000748 and 0.000784; both miss the 0.005 margin and have intervals crossing zero.Source: §A.2 · p. 15
• The separate projected-descriptor screen splits videos before standardization, projection, and router fitting, avoiding the preceding screen's feature-learning overlap.Both analyses use released computational descriptors rather than raw-media encodings.Source: §A.2 · p. 16

A.3 CONTROLLED PRE-QUERY VALUE SIMULATIONS

• In the simulator, cue-minus-random gain rises with cue accuracy, but an uninformative cue leaves matched-budget selection near random.With costs, unconstrained calibrated policies can request more often than matched-budget comparators; the simulator does not establish natural-source acquisition value.Source: §A.3 · p. 16

A.4 RAW-MEDIA SOURCE-FEASIBILITY STUDY

• The training-only Perception Test study fails its source-grounding criteria and does not support a source-grounded conflict benchmark under this design.The study reports 30–40% target accuracy, 40–55% donor accuracy, 15–25% pair support, and aligned gain of at most 10%.Source: §A.4 · p. 17, Table 5 · p. 17

A.5 CREMA-D NATURAL AUDIO–VIDEO EVIDENCE

• CREMA-D’s primary target used 7,022 singleton integrated votes from 91 actors assigned to task-fit, router-fit, calibration, and evaluation roles.The CREMA-D evaluation used 1,244 clips from 16 held-out actors; the frozen wav2vec 2.0–ViT comparison varied the number of uniformly sampled visual frames.Source: §A.5 · p. 17, Table 6 · p. 18
• For HuBERT–DeiT, audio–video accuracy was 56.43% versus 58.92% audio-only, a −2.49% difference with a 16-actor interval of [−8.40%, +2.81%].Its matched-request routing difference was also uncertain: +0.42% [−1.02%, +1.87%].Source: §A.5 · p. 18
• For wav2vec 2.0–ViT, eight-frame fusion reached 50.96% accuracy versus 45.10% for audio alone, but the matched-random difference was +0.04% [−0.30%, +0.36%].The actor intervals for the frozen-pair full-fusion and matched-random differences spanned zero.Source: §A.5 · p. 18, Table 7 · p. 19
• Across six overlapping actor reallocations, wav2vec 2.0–ViT had the largest descriptive margin at each calibration target, while the three swapped pairs showed no coherent sign pattern.These reallocations describe split sensitivity, not independent-population estimates.Source: §A.5 · p. 18, Table 8 · p. 19, Table 9 · p. 19

A.6 AMI AUDIO–VIDEO–TEXT PILOT

• On 244 donor-matchable AMI records, accuracy was 60.62% for audio, 44.60% for audio–video, 76.18% for text, and 35.77% for audio–video–text; routing text lagged static by 3.69%.The router requested text on 38 records (15.57%); at exactly 38 requests, its accuracy differences from random and confidence selection were −1.69% and −4.96%.Source: §A.6 · p. 20, Table 10 · p. 20, Table 11 · p. 20
• The AMI study was exploratory: only two meeting teams contributed evaluation cases, and the results did not establish a natural-source acquisition gain.Its fixed policy request rates made a common-rate comparison of static, all-request, and selective policies infeasible.Source: §A.6 · p. 20

B MATCHED SYNTHETIC CONTROLS

• The active and always-query-selective policies made identical decisions at different request rates, while query-majority kept the active request rate but had lower utility.The matched controls isolate request savings from the final-action rule; they do not show that the initial pair predicts when the third score changes an answer.Source: Table 12 · p. 21, B MATCHED SYNTHETIC CONTROLS · p. 21

C MATCHED-BUDGET SELECTOR VALUE

• At 10% and 25% budgets, the train-only selector beat matched random in utility, but fell below no-query at 50% and 63.7%; pair uncertainty had higher utility at every budget.At the two higher budgets, the selector’s non-ambiguous binary accuracy was slightly higher than pair uncertainty, while targeted decision accuracy favored pair uncertainty.Source: C MATCHED-BUDGET SELECTOR VALUE · pp. 21–22, Table 13 · p. 22

D COST-SENSITIVITY DETAIL

• At campaign noise, increasing threshold τ from 0.35 to 0.7 changed utility from 0.484 to −0.039; at τ = 0.35, increasing noise to 0.44 gave 0.337.The reported shifts reflect score magnitudes and the weak-score ambiguity marker, not robustness to arbitrary rescaling.Source: D COST-SENSITIVITY DETAIL · p. 22
• As query cost λq rose, threshold-policy utility decreased, while majority utility stayed constant because it neither queried nor abstained.Figure 3 re-scores fixed actions; the policy was not refit for the alternative costs.Source: D COST-SENSITIVITY DETAIL · pp. 22–23, Figure 3 · p. 23
• At campaign costs (λq, λa) = (0.10, 0.20), threshold utility was 0.481 ± 0.014 versus 0.252 ± 0.016 for majority.At (0.40, 0.80), threshold utility was 0.192 ± 0.018, below majority’s unchanged 0.252 ± 0.016.Source: D COST-SENSITIVITY DETAIL · p. 22, Table 14 · p. 23

E EXTENDED LIMITATIONS

• The active policy receives a third score after a request without modeling acquisition latency, missingness, privacy, or temporal alignment; its scores are not waveforms, frames, or language representations.CREMA-D routing matches request rate only and does not measure decoding delay, monetary cost, or deployment availability.Source: E EXTENDED LIMITATIONS · p. 23
• The synthetic generator uses a fixed threshold and few mechanisms; invert and swap share one score-level transformation, which mechanism-label summaries count twice.The paper identifies distinct transformations as a direction for future studies.Source: E EXTENDED LIMITATIONS · p. 23
• The temporal extension is not a natural-media evaluation, and equal-information controls match the temporal networks; the original factorial therefore establishes neither temporal-model advantage nor transfer to raw modalities.The zero-mean marker removes information from time means while leaving a fixed waveform in full traces; shared time permutation preserves the separation.Source: E EXTENDED LIMITATIONS · pp. 23–24
• The learned baselines use the same synthetic family, and the active rule receives the third coordinate after a request while fixed baselines receive all three initially.A fair real-media comparison would need defined modality availability, request latency, preprocessing cost, and a preregistered encoder set.Source: E EXTENDED LIMITATIONS · p. 24
• The designated ambiguity mechanism encodes a specific safety preference: abstention is rewarded there but penalized under mixed corruption, and changing λq or λa changes the preferred operating point.The study treats cost sensitivity as a required stress test before interpreting a policy as preferable.Source: E EXTENDED LIMITATIONS · p. 24
• The descriptor study uses released features and does not resolve the score abstraction; the Perception Test pilot also failed its source-grounding criteria and minimum unseen-component threshold.The study gathers no new annotations or human evaluation and does not address perceptual quality, accessibility, clinical use, or deployment readiness.Source: E EXTENDED LIMITATIONS · p. 24