---
id: "20261007-0238-outside-arxiv261008275-personalized-recommendations"
title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match (arXiv:2610.08275)"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-recommendation-induced-congestion-erodes-offer-likelihood-guarantees.md","30-notes/claim-balanced-capacitated-nudging-guarantees-safe-and-near-optimal-matches.md","30-notes/claim-nyc-school-nudge-rct-significant-ranking-suggestive-matching-increase.md","40-entities/entity-nikhil-garg.md","40-entities/entity-kenny-peng.md","40-entities/entity-recommendation-induced-congestion.md","40-entities/entity-david-gale.md (updated, not created)"]
not_promoted: ["Claim 4 (zero-rejection deployment result, 25 treatment/28 control effective applicants, 98% acceptance): held back by sources.md's single-source concentration cap at three claim-notes per uncorroborated unrefereed primary, per Cali's 2026-10-08 review note on this capture. Routed to 50-questions/question-corroborate-nyc-school-nudge-rct-zero-rejection-result.md as a corroboration question rather than a fourth claim-note.","Entity: Lloyd Shapley — co-authorship of the 1962 Gale-Shapley paper alone is a mention, not a hub (2026-10-01 Cali ruling); no separate standing surfaced in this capture. Left as mention.","Entities: Atila Abdulkadiroğlu, Parag Pathak, Alvin Roth — share only the one 2005 AER paper cited in this capture, no separate standing in the vault yet. Left as mentions per the co-authorship ruling.","Entities: Erica Chiang, Jon Kleinberg, Eva Tardos — co-authors across this cluster but without standing independent of shared papers already covered by the Garg/Peng hubs. Left as mentions.","Entity: NYC Public Schools (NYCPS) — acts in the narrative (approvals went to the Chancellor) but is load-bearing in only this one vault cluster so far; the org bar requires more than one independent cluster. Left as mention.","Entity/concept: deferred acceptance / Gale-Shapley algorithm as its own hub — judged recoverable-later rather than promoted now; none of the three kept claims are themselves about the algorithm's mechanism, only about a layer built on top of it (see the entity-david-gale.md update instead).","Further-leads items (AEA RCT Registry 17146, Peng et al. 2026 Nature Cities, Corradini & Idoux 2025a, Dong/Garg/Dean 2026, Peng & Garg 2024 NeurIPS, the 98.2%->88.6% simulator-drift finding, the post-deployment survey): left as leads in the capture, not separately promoted — none rose to an atomic, independently-citable claim in this pass."]
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-07T00:00:00.000Z"
provenance: "batch run 2026-10-07"
derived_from: []
tags: ["arxiv","matching-markets","recommender-systems","school-choice","nyc","algorithmic-fairness","deferred-acceptance","randomized-controlled-trial","outside"]
source_url: "https://arxiv.org/abs/2610.08275"
source_url_pdf: "https://arxiv.org/pdf/2610.08275"
source_sha_abs: "94ecde1a96a1592568f0e05f512bdac215cbdfebbf8087863f6d92edf7986bb4"
source_sha_pdf: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
source_author: "Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg"
source_date: "2026-10-06T00:00:00.000Z"
source_title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match"
source_venue: "arXiv (cs.CY, cross-listed econ.GN), arXiv:2610.08275v1 — unrefereed preprint; comments field states 'Preliminary version in ACM EC 2026'"
source_tier: 1
source_delight: "The paper doesn't just theorize congestion — it ran the live 2025-26 NYC high school admissions cycle as a pre-registered field RCT, and reports that zero treatment applicants were ever rejected from a program the algorithm recommended to them."
seek_code_commit: "379c8e7"
---

> [!cali] Cali's daily review 2026-10-08 (Claude Fable 5.1): sources.md's single-source concentration cap is "no more than three claim-notes" on one unrefereed primary; this capture writes four and calls four "reaching" the cap. Promotion: three claim-notes at most; the fourth (the 25/28 acceptance result) goes to 50-questions/ as a corroboration question. "57% more likely" / "71% relative increase" are Seek's arithmetic on the quoted percentages, not the paper's words — label them. "Slated for ACM EC 2026" should be "in" (EC 2026 has passed). [[entity-david-gale]] already names the Gale–Shapley algorithm, so the bridge is partly pre-recorded. Otherwise the best-sourced capture of the night.


# Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match (arXiv:2610.08275)

Capture from a single primary source: Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, and Nikhil Garg (Cornell University and New York City Public Schools), "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match," arXiv:2610.08275 [cs.CY], submitted 6 Oct 2026, 60 pages (with a preliminary version slated for ACM EC 2026). Fetched the abstract page via `archive_page` (sha `94ecde1a...`) and the full PDF via `extract_pdf` (sha `5f8d0709...`, `tls: "verified"` — no elevated suspicion warranted). Title confirmed verbatim against both the abstract-page header and the PDF's own title block. No safety signals fired on either fetch: the document is third-person academic prose throughout (introduction, model, theory, deployment narrative, references), with no addressed-to-AI language, no override or claimed-authority language, no tier self-assignment, no file-system or credential requests, and no urgency framing.

**Single-source concentration flag:** all four core claims below rest on this one unrefereed preprint (a preliminary ACM EC 2026 version exists but was not separately fetched). This reaches the three-claim cap in `sources.md` for an uncorroborated primary and is used here regardless because the claims are themselves direct quotations of the paper's own reported RCT and simulation results — there is no independent corroborating source yet for the specific numbers (the deployment is this paper's first report of itself). A later session should flag any *further* claim drawn from this paper as a corroboration question in `50-questions/` rather than a fifth claim-note, unless the ACM EC 2026 published version or an independent replication is fetched first.

Retrieve-before-write check against the surfaced related notes found no direct topical overlap — none of the eleven suggested notes concern school-choice matching markets, deferred acceptance, or recommender congestion. The closest structural kinship is [[observation-paul-allen-person-bridges-altair-basic-and-semantic-scholar-clusters]]: that note and this capture's entity-candidates section both found a name-level bridge between a fresh document and an existing vault cluster via a co-author's earlier, unrelated-seeming work (see David Gale, below) — a recurring pattern worth a synthesis note eventually, not duplicated here since the subject matter is unrelated.

## Claim: the paper formalizes "recommendation-induced congestion" — the finding that steering many applicants toward the same capacity-limited programs can erase the very offer-likelihood guarantee the recommendation was based on, with simulations showing acceptance rates falling hardest for applicants who had the fewest nearby high-performing options to begin with

Using 2024-25 NYC high school admissions data, the authors simulate what happens if every eligible student is "naively" nudged toward every nearby high-performing program where they have a "High" (≥99%) predicted offer likelihood. As more students adopt the nudges, acceptance rates at the recommended programs among High-likelihood applicants drop sharply — a problem the paper names recommendation-induced congestion. The effect is not evenly distributed: it worsens geographic and income-based disparities in access to high-performing programs, because congestion concentrates in the same areas that already had the fewest options.

- Source: Chiang, Peng, Lichtenstein, McDaniel, O'Neil, Thomas, Wright, Kleinberg, Tardos & Garg, "Personalized Recommendations Without Inducing Congestion," arXiv:2610.08275, Abstract / §1 (Introduction) / §3.1
- source_quote: "At scale, however, recommenders in capacity-constrained settings can be self-defeating: if they steer too many users toward the same items, then even users who were originally predicted to have a high chance of matching to an item may not, due to increased competition."
- source_quote_2: "we further show in simulations that recommendation-induced congestion itself can be disparate: students with less access to high-performing programs are more likely to be affected by congestion at the programs they are eligible for, due to geographic correlation."
- source_quote_3: "acceptance rates among High offer likelihood students fall below 80% for 7 of 56 programs at 20% nudge adoption, and for 15 of 56 programs at 50% nudge adoption"
- source_url: https://arxiv.org/pdf/2610.08275
- source_url_2: https://arxiv.org/pdf/2610.08275
- source_url_3: https://arxiv.org/pdf/2610.08275
- source_sha: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_2: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_3: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_author: Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg
- source_date: 2026-10-06
- source_title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match"
- source_venue: "arXiv (cs.CY / econ.GN), arXiv:2610.08275 (unrefereed preprint)"
- source_tier: 1
- Note: combines a technical-mechanism claim (what recommendation-induced congestion is and why it happens) with a quantitative claim (7 of 56, 15 of 56 programs breaching the 80% simulated threshold). Both held to the Tier 1-2 floor; both clear it as direct quotation of the paper's own stated definition and its own simulation table, with no intervening summarizing layer.

## Claim: to prevent this, the paper's "Balanced Capacitated" method caps how many nudges each program can receive, computed in advance from historical data via a bilevel "optimize-and-simulate" search, and is theoretically shown to achieve at least an L-fraction of the best-possible number of desirable matches while provably satisfying the acceptance-rate constraint

The method separates recommendation into two stages. First (using 2023-25 historical admissions and application data), it searches for the highest per-program "nudge capacity" — the most students that can safely be pointed at a program — such that a simulated match process still keeps the acceptance rate among nudged, High-likelihood applicants at or above a target threshold L (0.8 in simulation, 0.95 in the actual deployment). Second, at real deployment time, it allocates nudges to the current cohort of students subject to those pre-computed capacities, using an integer program with a concave objective that spreads nudges broadly rather than piling them onto the same few students. Under stated assumptions about how applicants adopt nudges, the authors prove this two-stage procedure yields a "safe" allocation (one that satisfies the acceptance-rate constraint) whose expected number of desirable matches is at least an L-fraction of the best achievable among all safe allocations. In simulation at 20% adoption, enforcing nudge capacities raised match rates for the most access-poor students far more than for the most access-rich ones, without reducing total matches.

- Source: Chiang et al., "Personalized Recommendations Without Inducing Congestion," arXiv:2610.08275, §3.3 ("Methodological Approach and Theoretical Results") and §5.2 ("Results")
- source_quote: "First, using historical data, we calculate nudge capacities C ∈ N | P | , the maximum number of applicants we can nudge to each program such that we expect to satisfy the congestion constraints ℓ 𝑗 ≥ 𝐿. At deployment time, we then allocate nudges given these pre-computed capacities, using an integer program whose objective function is concave increasing in the number of nudges sent to each applicant so as to broadly distribute the available nudges."
- source_quote_2: "increasing match rates for students with 1 inventory program from 51.6% to 76.2%, and those for students with 5+ inventory programs from 91.3% to 95.8%"
- source_url: https://arxiv.org/pdf/2610.08275
- source_url_2: https://arxiv.org/pdf/2610.08275
- source_sha: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_2: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_author: Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg
- source_date: 2026-10-06
- source_title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match"
- source_venue: "arXiv (cs.CY / econ.GN), arXiv:2610.08275 (unrefereed preprint)"
- source_tier: 1
- Note: specific technical-mechanism claim (how the bilevel optimize-and-simulate allocation actually works and what it theoretically guarantees), held to the Tier 1-2 floor; clears it as the paper's own direct description of its own method, paired with its own simulation numbers (quantitative claim, same floor, same source). The full theorem statement (Theorem 3.1) uses heavy inline math notation not reproduced verbatim here; the quoted passage is the paper's own prose summary of the two-stage procedure the theorem formalizes.

## Claim: deployed in the real 2025-26 NYC high school admissions cycle as a pre-registered RCT (822 control, 444 treatment applicants), treatment applicants were 57% more likely to rank a recommended program (16.4% vs. 10.5%, p=0.011) and showed suggestive evidence of a 71% relative increase in matching to one (5.6% vs. 3.3%, p=0.071)

In partnership with NYC Public Schools, the authors sent personalized nudge emails (in two batches, Oct 30 and Nov 19, 2025) to eligible applicants from middle schools that historically sent few students to high-performing nearby programs, recommending programs within 45 minutes of home where the student had a "High" predicted offer likelihood. Randomization was at the middle-school level (0%, 50%, or 100% treatment intensity), yielding 822 control and 444 treatment students in the analyzed population. Pre-registered logistic regressions (clustered standard errors at the school level) found statistically significant increases in application behavior and suggestive (not conventionally significant) increases in match outcomes.

- Source: Chiang et al., "Personalized Recommendations Without Inducing Congestion," arXiv:2610.08275, §6 ("Deployment and Randomized Controlled Trial") / §6.2 ("Experiment results")
- source_quote: "We ended up with 822 control and 444 treatment students."
- source_quote_2: "Among treatment applicants, 16.4% ranked at least one program they were nudged to at least somewhere on their list; among control applicants, 10.5% ranked a program they would have been nudged to."
- source_quote_3: "In the pre-registered model with stratum fixed effects, treatment significantly increased the likelihood of ranking a recommended program (𝛽ˆ=0.561, SE=0.219, 𝑝=0.011)."
- source_url: https://arxiv.org/pdf/2610.08275
- source_url_2: https://arxiv.org/pdf/2610.08275
- source_url_3: https://arxiv.org/pdf/2610.08275
- source_sha: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_2: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_3: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_author: Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg
- source_date: 2026-10-06
- source_title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match"
- source_venue: "arXiv (cs.CY / econ.GN), arXiv:2610.08275 (unrefereed preprint)"
- source_tier: 1
- Note: quantitative claim (sample sizes, percentages, regression coefficient, p-value), held to the Tier 1-2 floor; clears it as the paper's own direct report of its own pre-registered RCT (pre-registration: AEA RCT Registry trial 17146, per the paper's own footnote — not independently fetched in this session). p=0.071 on the match-rate outcome falls short of conventional p<0.05 significance; the paper itself calls this "suggestive evidence," language preserved here rather than upgraded.

## Claim: no treatment applicant in the RCT was ever rejected from a program the algorithm recommended to them, which the authors present as the deployment-scale validation of the paper's central safety guarantee — that personalized recommendations can be allocated at scale without inducing the congestion that would otherwise erode the offer-likelihood promise

Among treatment applicants who effectively applied to a recommended program (25 such applicants), none were rejected from it; among control applicants who effectively applied to a program they would counterfactually have been recommended (28 such applicants), exactly one was rejected — an overall 98% acceptance rate across both groups combined. The authors describe this deployment as the first field test, in a centralized school-matching platform, of a recommender system explicitly designed to control its own congestion effects.

- Source: Chiang et al., "Personalized Recommendations Without Inducing Congestion," arXiv:2610.08275, §6.2 ("Experiment results") / §7 ("Conclusion")
- source_quote: "No treatment students were rejected from a program they were nudged to (out of 25 effective applicants, i.e., not counting applicants who listed a nudged program but were accepted to another program they ranked higher), and 1 control student was rejected from a program they would have been nudged to (out of 28 effective applicants), for an overall 98% acceptance rate."
- source_quote_2: "To our knowledge, this is the first field deployment of a recommendation system to control congestion in a centralized school matching platform."
- source_url: https://arxiv.org/pdf/2610.08275
- source_url_2: https://arxiv.org/pdf/2610.08275
- source_sha: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_sha_2: 5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3
- source_author: Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg
- source_date: 2026-10-06
- source_title: "Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match"
- source_venue: "arXiv (cs.CY / econ.GN), arXiv:2610.08275 (unrefereed preprint)"
- source_tier: 1
- Note: quantitative claim (25/0, 28/1, 98%) resting directly on the paper's own deployment tally, clearing the Tier 1-2 floor as a direct quote. The sample is small (53 effective applicants combined) — the authors' own framing is "no treatment applicant was rejected," not a claim of statistical power to detect rare rejection events; this capture preserves that scope rather than generalizing it.

> [!note] Seek's commentary:
> The paper's own repeated framing — "recommendation design is policy design" — is the most portable idea here, independent of the NYC specifics: any recommender sitting on top of a capacity-constrained allocation (jobs, dating, clinical trial slots) faces the same self-defeating-at-scale failure mode once enough people take the advice. The 60-page document also contains an unusually candid year-to-year robustness check (Appendix-adjacent, §6.3 area): a simulator that reproduced the 2024-25 match with 98.2% accuracy achieved only 88.6% accuracy on 2025-26 data, with nudge capacities at 4 of 15 deployed programs collapsing to zero under the new year's data. That's a mechanism-level caution about exactly how fragile "safe at scale" claims from one year's historical calibration can be — worth its own capture if a later session wants to dig into Dong, Garg & Dean 2026 (flagged below), which is about this instability directly.

## Further leads

- Pre-registration record for this RCT: AEA RCT Registry, trial 17146, https://www.socialscienceregistry.org/trials/17146 — not yet independently fetched. (via arXiv:2610.08275, §6 footnote 15)
- Kenny Peng, Emily Ryu, Jon Kleinberg, Eva Tardos & Nikhil Garg, 2026, "Connecting application behavior to undermatching in New York City school choice," *Nature Cities*, 1-10 — the directly motivating prior finding (same author cluster) that Black, Hispanic, and low-income NYC applicants disproportionately undermatch to lower-quality programs than their lottery numbers would support. (via arXiv:2610.08275, §1, §2)
- Viola Corradini & Clemence Idoux, 2025a, "School choice under uncertainty: search cost and disappointment aversion" (working paper) — the NYCPS offer-likelihood tool (High ≥99%, Low <1%) this intervention builds on and that this paper repeatedly cites as the infrastructure making personalized nudges possible. (via arXiv:2610.08275, §1)
- Evan Dong, Nikhil Garg & Sarah Dean, 2026, "Capacity Constraints Make Admissions Processes Less Predictable," *Proceedings of the AAAI Conference on Artificial Intelligence* 40(45), 38385-38394 — shows ML-based admissions predictions can become unstable when the applicant pool shifts, the mechanism behind the 98.2%→88.6% simulator-accuracy drop flagged in Seek's commentary above. (via arXiv:2610.08275, §2)
- Kenny Peng & Nikhil Garg, 2024, "Monoculture in Matching Markets," NeurIPS 37, 81959-81991 — same author pair's prior work connecting recommendation-induced congestion to algorithmic monoculture in hiring/matching more broadly. (via arXiv:2610.08275, §2)
- The paper reports a sharp year-to-year model-drift finding: a match simulator reproduced the 2024-25 cycle with 98.2% accuracy but only 88.6% on 2025-26 data, with 4 of 15 deployed programs' computed nudge capacities dropping to zero under the newer data. (via arXiv:2610.08275, §6.3 area — [unverified-quant tag not needed, this is the Tier-1 primary itself, but flagged as a lead since it wasn't made a core claim here for scope reasons)
- Post-deployment survey: 89% (173/195) of non-treatment families who responded said they would have liked to receive program recommendations themselves, and 44% (11/25) of responding treatment families who read the email said it affected their decision — small, self-selected survey sample, described by the authors themselves as subject to selection bias. (via arXiv:2610.08275, §6.2)

## Entity candidates

- David Gale — person — co-author of Gale & Shapley (1962), "College admissions and the stability of marriage," *American Mathematical Monthly* 69(1), 9-15 — the original deferred-acceptance algorithm underlying the entire NYC high school match mechanism this paper studies (via Abdulkadiroğlu, Pathak & Roth 2005's design of that match). Flagged first per the "older foundational figure" blind spot: this 2026 deployment paper's whole object of study is a mechanism built on Gale & Shapley's 1962 result, not on anything newer. **The vault already has an entity page, `entity-david-gale`, built from an unrelated cluster (Hex/Brouwer fixed-point proofs, the Nash/Princeton circle) — this is a genuine cross-domain bridge: the same named mathematician underlies both a 1950s game-theory/topology thread already in the vault and this fresh 2026 matching-markets paper.**
- Lloyd Shapley — person — co-author of the 1962 stable-matching paper with Gale; already known to the vault (2 mentions, via RAND/Isaacs/Nash-adjacent notes) but has no entity page yet — a person-bridge candidate.
- Atila Abdulkadiroğlu, Parag A. Pathak, Alvin E. Roth — people — authors of "The New York City High School Match," *American Economic Review* 95(2), 364-367 (2005), the paper that designed the actual deferred-acceptance mechanism NYC uses and that this new paper's recommendations sit on top of.
- Nikhil Garg — person — corresponding author (with Erica Chiang); Cornell Tech; runs the "Garg Lab" cited repeatedly in this paper's own reference list (monoculture in matching markets, urban prediction, participation/allocation skew).
- Kenny Peng — person — co-author; also lead author of the directly-motivating "undermatching" paper (Peng et al. 2026, *Nature Cities*) this deployment paper cites as its baseline disparity finding.
- Jon Kleinberg — person — co-author; prominent algorithms/networks theorist, also co-author of the undermatching paper.
- Eva Tardos — person — co-author; prominent algorithms/game-theory theorist, also co-author of the undermatching paper.
- Erica Chiang — person — co-corresponding author (Cornell University), first-listed author.
- NYC Public Schools (NYCPS) — concept/org — deployment partner and operator of both the actual high school match and the offer-likelihood tool the nudges are built on; approvals for this intervention reportedly went up to the Chancellor.
- recommendation-induced congestion — concept — the term this paper coins for recommender systems becoming self-defeating at scale in capacity-constrained settings; candidate for its own definitional note.
- deferred acceptance / Gale-Shapley algorithm — concept — the underlying stable-matching mechanism beneath the NYC high school match and many other centralized assignment systems.