acceptodds
Under review as a conference paper at ICLR 2027

A Correct Mask Hides the Source of Failure: Two Kinds of Observation Aliasing in Learned Routing

Abstract

Learned routing policies act under a correct feasibility mask, yet can still rank feasible actions poorly. We show that such a quality failure has more than one source, and that a correct mask hides which one: the deficit depends on where the decision-relevant information that the decoder observation aliases resides. In controlled experiments with a construction policy implemented independently of any released system, we exhibit a dissociation. When the aliased quantity is a state that accumulates along the route, adding decoder memory reconstructs it and removes 95% of the deficit over fifty-two training seeds. Two comparisons place that recovery: a content-free channel of the same width removes nothing, while the same recurrent module fed a constant removes 48%, mostly because its state is zeroed at each depot visit. Driving that reset from a different instance's depot schedule, matched to about a percent in how deep the state is when it is read, removes nothing (to within 0.07): what the reset supplies is alignment with the instance's own tour, not re-anchoring at the right rate. About half the recovery comes from depth into the tour and half from the route's accumulated state. When the aliased quantity is a per-instance latent regime that leaves no signature in the observed stream, decoder memory cannot reconstruct it and does not help; only an explicit regime signal closes the deficit. The same added memory therefore repairs one failure and leaves the other essentially unchanged, and the separation holds across regime strengths in the latent case and across model widths in the reconstructible case. We then revisit a released multi-task routing model whose mixed-backhaul failure combines both ingredients, an unseen regime and an aliased feasibility state, and where, after fine-tuning on a pool that covers the regime, a one-bit regime signal raises the repair of its 6.2 pp zero-shot deficit from 0.47 to 0.88. In practice, therefore, the source must be diagnosed before a fix is chosen. The contribution is a mechanism-level dissociation: two hidden sources of failure that a correct mask renders indistinguishable, and that demand different fixes.

Primary area:
reinforcement learning
Keywords:
neural combinatorial optimization, vehicle routing, observation aliasing, partial observability, recurrent memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r9GEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the vehicle routing problem, where a sequence of actions corresponds to a sequence of edges connecting a graph. In vehicle routing tasks, the action space is often masked because the roads will not allow the vehicle to go in any direction at any time. The authors posit that when learning to navigate this graph fails, the mask hides the cause of the learning failure.

The authors train an attention-based autoencoder using REINFORCE across two variants. From what I understand, one variant is fully observable and one is partially observable. Both approaches result in a "quality deficit". Adding a recurrent model (memory) seems to improve results.

Critical strengths

There might be some nuggets of useful information here hidden beneath AI-prose, but I struggle to find them.

Critical weaknesses

This paper has undergone heavy AI-editing that makes it very difficult for human readers to understand. Consider the first two sentences of the abstract:

Learned routing policies act under a correct feasibility mask, yet can still rank feasible actions poorly. We show that such a quality failure has more than one source, and that a correct mask hides which one: the deficit depends on where the decision-relevant information that the decoder observation aliases resides.

It is unclear what the policies are "routing" (e.g., LLM, MoE?), what a "feasibility mask" is, what it means for the mask to be "correct", what a "quality failure" is, and so on. Of course, these are just the first two sentences of the abstract. The problem repeats itself throughout the entirety of the paper.

Beyond the dense writing, the contribution and novelty is unclear. It seems like the authors found that memory (recurrence) helps to solve POMDPs, and somehow associated this with action masking. It is well known that memory provides a solution to POMDPs.

Questions for authors
  • What does it mean for an autoencoder to be trained using REINFORCE? A REINFORCE objective uses the return but an autoencoder objective uses reconstruction.
Reviewer_2chpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates poor solution quality in neural routing policies operating under correct feasibility masks. It distinguishes missing quantities reconstructible from the route from per-instance latent regimes absent from the observable history. Controlled routing tasks compare direct information access, recurrent summaries, and content-free controls. Route memory recovers approximately 95% of the Family A deficit relative to direct access, but offers essentially no recovery in Family B. Further experiments examine depot resets and partially informative observations. A RouteFinder mixed-backhaul case study examines training coverage and compares candidate input signals on a common fine-tuning pool.

Critical strengths
  • The controlled comparisons are carefully designed. The main experiments use many training seeds and distinguish direct information access, static content-free channels, and recurrent controls. Additional width, training-budget, scale, and evidence-noise experiments support the reported dissociation within the tested settings, providing substantially stronger evidence than a simple comparison with and without memory.
Critical weaknesses
  • Since the organizing distinction follows established partial-observability principles, the practical contribution depends on the experiments. These carefully validate tasks with known information structure, but do not establish how reliably Section 7's checks distinguish unavailable information from failed recurrent learning in an unknown system. Broader diagnostic reliability remains insufficiently established.
  • The released-model case provides limited validation of the two mechanisms beyond the controlled tasks. Unlike Family B, its regime changes the mask, leaving inferability from history unresolved. Content-fed and noise-fed recurrence are statistically indistinguishable, and the backbones share the relevant implementation. These limitations constrain mechanism attribution and generalization beyond this ecosystem.
Questions for authors
  • What evidence distinguishes a regime unrecoverable from history from a reconstructible quantity that the recurrent policy fails to learn when applying Section 7's checks? What conclusions are justified when route content provides no measurable benefit over recurrent noise input?
  • Can RouteFinder's class-2 regime be inferred from the complete observation–action–mask history? If so, how should the relationship between this case study and Family B's history-independent regime be interpreted?
  • Given the statistically indistinguishable content-fed and noise-fed recurrent results in Section 6, what do the separate-load and coupled-headroom interventions establish about trajectory-reconstructible state aliasing beyond a local decoder-context mismatch?
  • For the RouteFinder recurrent experiments, what are the recurrent inputs and architecture, and which base-context and constant-bit runs serve as controls? Can you provide the corresponding implementation and specify the common pool's size and generation procedure?
Reviewer_CbpsOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies observation aliasing in learned routing under a correct feasibility mask. It distinguishes missing state information that can be reconstructed from the routing history from latent regimes that leave no observable trace. Controlled experiments compare recurrent memory with explicit access to the missing information. A RouteFinder case study further examines the effects of training coverage, load information, and a regime signal.

Critical strengths

The method is partly clear.

Critical weaknesses
  1. The main conclusion appears largely built into the task design: memory can help when history contains the missing information, but cannot recover an independent hidden regime. What substantive insight does this study provide beyond established partial-observability arguments?

  2. The role of the correct mask in the contribution is unclear. Feasibility does not guarantee good action ranking. What new understanding of learned routing does the proposed distinction establish beyond the consequences of incomplete policy inputs?

  3. The practical recommendation is to diagnose the missing information before choosing memory or an explicit signal. However, the controlled tasks specify this information in advance. How does the study help people identify the failure source when it is actually unknown?

Questions for authors

See above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.