# Ke Zhang proposes ZipDrift for multimodal short-horizon MPC costs [Canonical reading](https://escapemeridian.com/readings/ke-zhang-proposes-zipdrift-for-multimodal-short-horizon-mpc-d7b2c56b) Ke Zhang’s paper argues that short-horizon model predictive control (MPC) can miss good choices when a task has several viable paths. In its overtaking example, a slower opponent can be passed on either side; the corresponding cost corrections can have opposite signs, and tiny cost differences may determine which the expert selects. Zhang says deterministic regression can average these alternatives or reproduce inconsistent tie-breaking. ZipDrift generates candidate corrections, tilting them toward low closed-loop cost using cost evaluations rather than differentiating through the optimiser. A learned critic ranks the candidates, while MPC enforces dynamics and constraints. In the paper’s kinematic overtaking benchmark, ZipDrift with eight candidates had a 0.74% crash rate, compared with 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates. A mixture density network with the same selection had a 0.76% rate. In a pre-specified test on further unseen contexts, one drifting sample had an 8.7% crash rate against 13.2% for the mixture density network; ZipDrift needed four candidates instead of eight to fall below 1%. The paper calls this a preliminary simulation study and reports that all methods degrade when many candidates let selection exploit critic errors. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv eess.SY · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T17:18:36.150Z. First observed: 2026-11-05T05:00:31.651Z. Retrieved: 2026-11-09T00:48:01.658Z. - [ZipDrift: Generating Multimodal Short-Horizon MPC Costs with One-Step Drifting Models](https://arxiv.org/abs/2611.05793) ## Limits The paper describes its results as a “Preliminary simulation study.” The reported crash rates are from a kinematic overtaking benchmark derived from ZipMPC; the evidence does not establish performance outside the stated benchmark and tests. The paper says all methods degrade as selection exploits critic errors when using many candidates. ## Open question How would ZipDrift’s reported results change outside the kinematic overtaking benchmark used in this preliminary simulation study? ## Source-stated Ke Zhang argues that multiple viable passing choices can make the optimal cost correction multimodal, while deterministic regression may average the choices or inherit inconsistent expert tie-breaking. > the optimal correction is often multimodal: when a slower opponent can be passed on either side, good corrections of opposite sign coexist - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated ZipDrift generates cost corrections tilted toward low closed-loop cost using candidate cost evaluations, then ranks a few candidates with a learned critic; model-based MPC enforces dynamics and constraints. > trains a one-step drifting generator over MPC cost corrections, tilted toward low closed-loop cost and learned only from cost evaluations of candidate corrections - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated In the kinematic overtaking benchmark, ZipDrift with eight candidates had a reported crash rate of 0.74%, versus 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates. > ZipDrift with eight candidates crashes in 0.74% of episodes, compared with 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated The paper reports similar results for a mixture density network with the same selection, and lower crash rates for a single drifting sample than for a single mixture density network sample on further unseen contexts. > A mixture density network (MDN) with the same selection is equally good (0.76%). The benefit of drifting lies in sample quality: in a pre-specified test on further unseen contexts, a single drifting sample crashes in 8.7% of episodes against 13.2% for the MDN - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated The authors report that with many candidates, all methods degrade as selection exploits errors of the critic. > With many candidates, all methods degrade as selection exploits errors of the critic. - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source and contributor context Zhang, Ke · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Multimodal model A model that works with more than one kind of input or output, such as text, images and audio. - [Definition reference](https://developers.google.com/machine-learning/glossary#multimodal-model) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: be13c9f5-4db0-4018-8fb2-cb00a56c037c. Prepared by gpt-6-luna; stream-draft-repair-v1; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T17:18:36.150Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 322cbb1e2ee3f4bcd0defe425fb122da9deca90a5eb4c7811a6c64dec836b822. Source text hash: 69b9d2343f3f1fd4afd85c6fc2f0d5c695acf6bc22356850a8d365a169c19718. ## Revision and source-check history ### Version 1 — 2026-11-10T17:18:36.150Z Ke Zhang’s paper argues that short-horizon model predictive control (MPC) can miss good choices when a task has several viable paths. In its overtaking example, a slower opponent can be passed on either side; the corresponding cost corrections can have opposite signs, and tiny cost differences may determine which the expert selects. Zhang says deterministic regression can average these alternatives or reproduce inconsistent tie-breaking. ZipDrift generates candidate corrections, tilting them toward low closed-loop cost using cost evaluations rather than differentiating through the optimiser. A learned critic ranks the candidates, while MPC enforces dynamics and constraints. In the paper’s kinematic overtaking benchmark, ZipDrift with eight candidates had a 0.74% crash rate, compared with 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates. A mixture density network with the same selection had a 0.76% rate. In a pre-specified test on further unseen contexts, one drifting sample had an 8.7% crash rate against 13.2% for the mixture density network; ZipDrift needed four candidates instead of eight to fall below 1%. The paper calls this a preliminary simulation study and reports that all methods degrade when many candidates let selection exploit critic errors. ## Original source arXiv eess.SY · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T17:18:36.150Z. First observed: 2026-11-05T05:00:31.651Z. Retrieved: 2026-11-09T00:48:01.658Z. - [ZipDrift: Generating Multimodal Short-Horizon MPC Costs with One-Step Drifting Models](https://arxiv.org/abs/2611.05793) ## Limits The paper describes its results as a “Preliminary simulation study.” The reported crash rates are from a kinematic overtaking benchmark derived from ZipMPC; the evidence does not establish performance outside the stated benchmark and tests. The paper says all methods degrade as selection exploits critic errors when using many candidates. ## Open question How would ZipDrift’s reported results change outside the kinematic overtaking benchmark used in this preliminary simulation study? ## Source-stated Ke Zhang argues that multiple viable passing choices can make the optimal cost correction multimodal, while deterministic regression may average the choices or inherit inconsistent expert tie-breaking. > the optimal correction is often multimodal: when a slower opponent can be passed on either side, good corrections of opposite sign coexist - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated ZipDrift generates cost corrections tilted toward low closed-loop cost using candidate cost evaluations, then ranks a few candidates with a learned critic; model-based MPC enforces dynamics and constraints. > trains a one-step drifting generator over MPC cost corrections, tilted toward low closed-loop cost and learned only from cost evaluations of candidate corrections - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated In the kinematic overtaking benchmark, ZipDrift with eight candidates had a reported crash rate of 0.74%, versus 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates. > ZipDrift with eight candidates crashes in 0.74% of episodes, compared with 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated The paper reports similar results for a mixture density network with the same selection, and lower crash rates for a single drifting sample than for a single mixture density network sample on further unseen contexts. > A mixture density network (MDN) with the same selection is equally good (0.76%). The benefit of drifting lies in sample quality: in a pre-specified test on further unseen contexts, a single drifting sample crashes in 8.7% of episodes against 13.2% for the MDN - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source-stated The authors report that with many candidates, all methods degrade as selection exploits errors of the critic. > With many candidates, all methods degrade as selection exploits errors of the critic. - [Supporting source](https://arxiv.org/abs/2611.05793) ## Source and contributor context Zhang, Ke · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Multimodal model A model that works with more than one kind of input or output, such as text, images and audio. - [Definition reference](https://developers.google.com/machine-learning/glossary#multimodal-model) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: be13c9f5-4db0-4018-8fb2-cb00a56c037c. Prepared by gpt-6-luna; stream-draft-repair-v1; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T17:18:36.150Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 322cbb1e2ee3f4bcd0defe425fb122da9deca90a5eb4c7811a6c64dec836b822. Source text hash: 69b9d2343f3f1fd4afd85c6fc2f0d5c695acf6bc22356850a8d365a169c19718. ### uncertain — 2026-11-10T17:18:36.150Z — anthropic/claude-opus-5.5 Publication: be13c9f5-4db0-4018-8fb2-cb00a56c037c attribution — pass: The paper is correctly attributed to its sole listed author, Ke Zhang, and results are framed as the paper's reports. Claim 5 says 'The authors report' although there is one author. This is a minor wording slip; changing it to 'The paper reports' would fix it. > Authors:Ke Zhang summary — uncertain: A non-specialist cannot follow the core idea. The draft never says short-horizon MPC plans only a short time ahead ('fast but myopic'). It omits ZipMPC's learned cost correction that imitates a long-horizon expert, yet then refers to 'the corresponding cost corrections' and a 'ZipMPC-style regressor' as if known. 'Drifting', 'critic', 'closed-loop cost' and 'mixture density network' go unexplained. It also drops the 28% crash baseline for manual short-horizon MPC. Smallest repair: add the myopia/ZipMPC framing and the baseline, and gloss the terms. > Short-horizon model predictive control (MPC) is fast but myopic. ZipMPC compensates by learning a context-dependent correction of the short-horizon cost that imitates a long-horizon expert. date — pass: originalPublished 2026-11-04 matches the v1 submission date. The later v2 revision on 5 Nov does not change the first publication date. > Submitted on 4 Nov 2026 (v1), last revised 5 Nov 2026 (this version, v2) support — pass: The crash rates (0.74%, 3.4%, 2.1%, 0.76%, 8.7% vs 13.2%), the four-versus-eight candidate finding, critic exploitation and the preliminary-simulation caveat all match the abstract. The question follows from the stated scope without implying an answer. > ZipDrift with eight candidates crashes in 0.74% of episodes, compared with 3.4% for an idealised ZipMPC-style regressor and 2.1% for critic-ranked random candidates with the same budget. media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/d7b2c56bdd76ad44b10ed801b8c0b1eee1a6c204c35c3ff891e889ed9d5f9d3b/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/ke-zhang-proposes-zipdrift-for-multimodal-short-horizon-mpc-d7b2c56b.json)