Executive Summary

Key Figures

The first two are the two counting rules pulling apart under the same duplication. The last two are the admission record and the one task that record hides.

Source: arXiv 2609.01662 (2026-08-31) · 60 episodes from the HABIT dataset, 720 offline evaluations

The test bed is a dataset another team released. HABIT collects scenes of people and robots handing objects to each other: six tasks, 696 episodes, and synchronized footage from five cameras. Query each image under four different prompts and a single camera yields four answers from one viewpoint. The authors retrained nothing. They varied only how those outputs are grouped and counted.

The admission study used 60 episodes held apart from the 82 development episodes. Each episode contributes two time windows, each holding one true target and five distractors, which produces 720 verdicts. Taking camera identity as the provenance boundary, so that the four prompt variants of one image fall into one component, PACT admitted 47 of 57 candidates consistent with the reference action on Qwen3-VL-32B, and 43 of 53 on Qwen3-VL-8B. No admission inconsistent with the reference was observed at either checkpoint.

Then came the duplication. Outputs were copied within a single camera until the total multiplicity reached eight. The authors record the result in one line: "Within-camera duplication to total multiplicity eight leaves all 720 final typed responses per checkpoint unchanged." Repeating every selected output moved neither the evidence budget, nor the posterior, nor the temporal comparison that decides whether this is the moment to act.

Run the same manipulation through a counting rule that treats every output as its own unit and the picture changes. Under the binary evidence mapping, which assigns total mass two to each output, the evidence budget of one event window rose by as much as 280 units. The authors add that at this operating point the inflated budget still changed none of the event decisions. They are marking the limit of their own result: a budget that swells is not by itself a verdict that collapses.

Why not just detect the duplicates? Design the confidence score well, measure the conflict between outputs, attach a few quality metrics, and copies ought to fall out. The paper's first proposition takes that expectation head on.

Take the simplest possible case. There are two inputs, and each presents two identically valued entries. In the first input the second entry is a copy of the first and shares its parent. In the second input the second entry happens to carry the same value but comes from a separate origin. Any fusion map that reads only values receives the identical multiset in both cases and must therefore return the identical answer. The authors label this Proposition 1, value-only indistinguishability.

No amount of sophistication in the quality score solves this problem. However many outputs support a candidate, how many distinct source groups they came from has to be recorded separately. High local reliability and low conflict do not change that. Countability follows from group membership, not from values.

This trap is not one the paper discovered. The related work section names the prior formalizations directly: covariance intersection, which estimates conservatively when cross-covariances are unknown; data-incest methods, which use information-flow structure to prevent the same information from being reused; Denœux's cautious rule, which stays idempotent for reliable but potentially overlapping bodies of evidence; and Subjective Logic, which separates cumulative from averaging fusion for independent versus dependent opinions. The authors call these substantive precedents for preventing unsupported accumulation under dependence. What PACT adds on top is the supplied partition as an explicit variable, and an answer to what the largest evidence budget permitted under that partition is.

PACT fits in two lines. It receives, from outside, a partition declaring which outputs came from the same observation. Within a component it keeps only the coordinatewise minimum, which the paper calls the meet, and it sums only across components. However many copies you push into a component, the minimum does not grow, so duplication buys no budget.

This rule was tested on one more bench besides the robot data of the previous section. The authors built a simulation benchmark where they could assign provenance relations by hand. It holds 31,200 evaluations nested in 48 scene clusters, built from Isaac Sim scenes, with three sources: language, geometry, and path risk. Geometry and risk share a scene-context parent, so they form one component while language stands alone. This benchmark is what makes it possible to lay the same opinions over different partitions and compare. The 24,000 instances and the 96.4% below come from here as well.

A component with only one member has to keep its value, and inserting a new member must not raise the budget. Among the rules that satisfy both requirements at once, the authors show, the coordinatewise meet is the one granting the largest budget. So the minimum is not an arbitrary pick. They also ran the contrast, replacing the minimum with the coordinatewise maximum. Maximum survives exact copies but violates insertion non-amplification: it raised the component budget on 96.4% of the 24,000 instances where geometry and risk were observed and valid. The meet holds up even when a copy is not exact. Across 960 comparisons spanning four perturbation magnitudes, inserting a near copy changed no predicted contract, and the largest posterior drift was 0.0109.

Keeping one representative per component, the natural compromise, was tested too. The authors' finding is that a single representative does not generally recover the component meet. Using one designated source per component gave a pre-admission risk metric of 0.8639 against 0.6294 for PACT fusion. After admission the gap runs 0.1668 to 0.0861. At a target coverage of 0.13, the representative produced 246 wrong admissions among 4,051, while PACT had none among 4,043. Nothing guarantees that the value of the chosen representative equals the coordinatewise minimum of its component.

3.1PACT Holds, Confirms, or Falls Back Instead of Rejecting

Gathering the evidence does not authorize the action. PACT filters candidates once by score, then checks four conditions in order: command consistency, source validity, risk support, and component corroboration. The first condition that fails determines the type of the response, and the paper maps them in order to hold, fallback, hold, and confirm. Passing all four yields admission.

The last condition catches the simulation benchmark in an instructive way. The authors required support from three distinct complete components, while the benchmark's partition has only two, language in one and geometry with path risk in the other. Even when all three sources agree unanimously on the same action, they occupy two components, so the system requests confirmation rather than admitting. A single number carries the design: unanimity is not a pass.

On the real robot data, two more signals join the admission conditions. Target identity asks whether the object being handled is the right one, and event proximity asks whether this is the moment for the action. Neither is tallied like a vote; each is a requirement that has to be met on its own. The table below shows what happens on Qwen3-VL-8B as the signals are added one at a time.

Consistency, in the table above, means something narrower than it sounds. It means the decision matched the action the dataset designates as correct, not that the robot completed the task. The authors say as much: this study measures reference consistency rather than physical success, and task success, the response burden placed on the person, and perceived safety are left to closed-loop and behavioral studies.

Directly after the 47 admissions and zero false ones, the authors write that this aggregate result conceals a task boundary. On Hotel Towel, one of the six tasks, PACT admitted none of the ten reference-ready episodes at either checkpoint. Recall zero. Average across the tasks and this boundary disappears.

How missing sources are handled also swings the record hard. The paper's primary convention leaves an expected but unobserved source in place as a zero vector, and under that convention the pre-admission risk metric was 0.6294 and the post-admission one 0.0861. Three alternatives, which exclude the zero, remove the unavailable source, or move it into an isolated component, each yield 0.7748 and 0.1875. That second value is worse than the 0.1479 that nested Dirichlet obtains under its own missing-source convention. The authors report it themselves: one purely definitional choice can reverse a comparative advantage outright.

Even where every source is observed, the ordering is not fixed. Scored by PACT's own selection score, PACT leads. Rescored by a shared score built on the posterior peak, the distance to nested Dirichlet narrows and plain product fusion gains ground. How outputs are grouped and counted, and what they are ranked by, are decisions made in different places, and being good at one does not carry over to the other.

The authors also measured how much of the work the admission conditions actually do. They planted 2,400 cases of high-confidence but incorrect agreement inside the benchmark as a stress test. Turn the component corroboration condition off and all 2,400 are admitted; turn it on and all of them route to a confirmation request. Yet on the remaining 28,800 instances, with those 2,400 excluded, toggling corroboration changed the risk metric not at all. Grouping the sources by provenance, over that same set, lowered the metric from 0.1250 to 0.0693. Counting by group and deciding how many groups must back a candidate do different jobs. The second is not a device for raising everyday performance but one that fixes the response to a trap, and the authors note that the trap arm is constructed by design.

Rather than leave this warning at the level of words, the authors corrupted the partition deliberately and measured the price. With half the relations wrong, false splitting and incorrect distinct-parent assignments produced an overall error rate of 0.0492, which is 43.8% below the 0.0875 of a provenance-unaware reference that counts every output singly. Push the corruption to 100%, the maximum tested level, and that advantage disappears exactly, landing on the reference. A wrong partition beats no partition up to a point, and past that point nothing is left of the gain.

The authors wrote this extension into the paper's discussion themselves. They cite self-consistency, which samples multiple reasoning paths from one input, alongside work showing that related language models produce correlated errors, and then write that when outputs share a derivation, output multiplicity alone should not be treated as separately countable support. The distinction their experiments draw, prompt variants of one image as repeated inference and distinct synchronized cameras as separate acquisitions, is one they carry straight over to language models.

They also ran the comparison that groups checkpoints by model family instead of by camera. Across a panel of eight vision-language checkpoints they checked whether family labels capture error association, and in this panel neither the family labels nor output similarity established countability. That is not the same as saying family grouping was a poor candidate. Ranked by within-minus-between error association, the model-family pairing came seventh among 105 pair partitions, with nine pairings tied with it or ahead of it. Weight the tasks equally and it moves to 22nd, with 24 tied or ahead. The accurate statement is that the data does not single out the family partition as the answer.

The opposite mistake has a price of its own. Merging all eight checkpoints into one component retains 9.9% of the budget held under model-family grouping and drops accuracy from 86.6% to 76.3%. Model-family grouping, the authors conclude, is a counting choice to be tested rather than inferred from architecture. Throughout, the partition was supplied explicitly and then checked against the observed dependencies.

The Pebblous blog recently covered an experiment where LLMs read the same papers five times and wavered. That study reported that pooling answers from different models agreed with human reviewers better than running one model more often. This paper adds a condition on top. Changing the model does not by itself make an observation separate. What counts as a distinct source still has to be declared by a person and checked against the data.

Thank you for reading. The paper and the code the authors released are at arXiv 2609.01662 and in the GitHub repository. The affiliation list names five universities including McGill and UBC, plus Mila – Quebec AI Institute and NVIDIA, and the simulation benchmark was built in NVIDIA Isaac Sim.

Pebblous Data Communication Team
September 4, 2026

Primary Paper, Dataset, and Code

• 1.Jin, Z., Zhang, H., Tang, Y., Hu, F., Dong, Z., & Shao, Y. (2026). "Not All Agreement Counts as Corroboration: Provenance-Conserving Multi-View Fusion for Typed Action Admission in Human–Robot Collaboration." arXiv:2609.01662.
• 2.Song, J., Jeong, S., Jeon, B., Kim, S., Seo, M., Son, H., & Lee, K. (2026). "HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation." arXiv:2606.31682.
• 3.Jin, Z. (2026). PACT official repository — code, experimental settings, and evaluation resources.

Prior Formalizations and Related Work

• 4.Denœux, T. (2008). "Conjunctive and Disjunctive Combination of Belief Functions Induced by Non-Distinct Bodies of Evidence." Artificial Intelligence, 172(2–3), 234–264.
• 5.Jøsang, A., Diaz, J., & Rifqi, M. (2010). "Cumulative and Averaging Fusion of Beliefs." Information Fusion, 11(2), 192–200.
• 6.Wang, X., Wei, J., Schuurmans, D., Le, Q. V., Chi, E. H., Narang, S., Chowdhery, A., & Zhou, D. (2023). "Self-Consistency Improves Chain of Thought Reasoning in Language Models." The Eleventh International Conference on Learning Representations (ICLR).
• 7.Kim, E. M., Garg, A., Peng, K., & Garg, N. (2025). "Correlated Errors in Large Language Models." Proceedings of the 42nd International Conference on Machine Learning (ICML), 267, 30038–30066.