{"id":"3e8150b3-dbfc-41d4-8638-ae87743d0a28","arxiv_id":"2607.14859","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"In a simulated two-route city, randomized CAV routing beats system-optimal routing when fleets chase market share, but adding average travel time into the objective reverses that advantage.","lead":"This paper simulates competing fleets of autonomous vehicles that route drivers to maximize market share, and finds that randomized, unpredictable routing can win more customers than standard system-optimal routing—at the cost of unstable travel times. It then proposes adding average travel time to fleet incentives to discourage such antisocial randomization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Randomization's causal mechanism is absent from the model: HDV utility uses only average travel times, so 'unpredictable travel times' cannot drive the reported advantage.","rationale":"The Reader's weakest assumption (known discount factors) is real but is explicitly declared as an assumption in Table 1 and Section 3, with inference listed as future work. A more foundational gap is internal: the mechanism advertised in the abstract—unpredictable travel times for HDVs inducing switching—is not represented in the utility specification. Eq. (2) defines HDV utility over average times only; route choice also uses average times. Therefore any day-to-day variation in travel times caused by random CAV routing has no channel through which to affect HDV mode choice, except through changing averages. A deterministic schedule with the same frequencies yields the same averages and identical linear credibility updates, so the reported RFlex advantage likely survives without randomness. This does not disprove the benchmark result, but it undermines the causal narrative and the Experiment 3 policy interpretation. A simple control experiment (deterministic equivalent) would settle it. Hence I recommend retaining the CONDITIONAL verdict but requiring this control and a corresponding text revision.","tokens_in":8631,"tokens_out":8252,"duration_ms":84515,"concrete_test":"Run the Experiment 2 one-fleet and two-fleet scenarios with RFlex/RFlexV modified to use a fixed, publicly announced periodic faster-route schedule (same per-driver frequencies as the random algorithm, e.g., deterministic round-robin), keeping everything else identical, over at least 20 seeds. If market share and mean travel time trajectories are statistically indistinguishable from the randomized version, randomization is not the operative mechanism and the abstract/introduction must be reworded; if the advantage disappears, the model does encode unpredictability and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes the efficiency of RFlex/RFlexV to 'unpredictable travel times for HDVs' (Abstract; Section 1, item iii). But the HDV model in Section 2 cannot respond to unpredictability: u_HDV is a logit-weighted average of past route travel times (Eq. 2), and HDV route choice uses the same logit on mean times; no variance, risk-aversion, or day-specific information enters. Credibility updates (Eq. 4) are linear in realized O/t, so a deterministic turn-taking schedule with identical per-driver frequencies would produce the same expected credibility and the same mode choices. The observed advantage of RFlex over SO may therefore be due to selective assignment of high-gamma drivers to the faster route and to kappa=0.5 offers, not to randomization per se. Because the abstract and the Experiment 3 policy discussion target 'antisocial randomised strategies', the model does not test the stated mechanism. This is load-bearing for the title/abstract claim, even if the benchmark result itself stands.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a simulated market in which human drivers (HDVs) choose between driving independently and joining one of two competing CAV fleets, with fleet revenue proportional to market share. The authors propose several routing algorithms: proportional system-optimum (SO) and user-equilibrium (UE) baselines, and two randomized heuristics (RFlexV and RFlex) that route high-discount-factor drivers to the faster route while randomizing assignments. In a one-OD-pair, two-route BPR network, they report that randomized algorithms achieve higher market share than SO/UE when drivers' discount factors are known and diverse. They also propose a combined objective Obj_mu = (1-mu) n_f + mu tau_avg and show that for mu approx 0.5 the SO-based algorithm becomes competitive, which they interpret as discouraging antisocial randomization. The paper claims that randomized routing works by making travel times unpredictable for HDVs, and that a socially oriented objective preserves competition while limiting this behavior.","tokens_in":8954,"tokens_out":3390,"duration_ms":33871,"significance":"If the central claim held, the paper would make a useful contribution to the design of future CAV routing markets: it would show that market-share-maximizing fleets can exploit randomized routing to attract reluctant HDV drivers, and that a simple modification of the fleet objective can mitigate the resulting system inefficiency. The benchmark framework and explicit algorithm descriptions are valuable, and the SUMO microsimulation appendix is a positive step toward external validity. However, the significance is substantially weakened by the mismatch between the stated causal mechanism and the model, the mixed results in two-fleet scenarios, and the lack of statistical rigor (no error bars, no sensitivity analysis, parameter choices made after inspecting results). The paper introduces a useful starting point, but the evidence as presented does not support the strength of the abstract's claims.","major_comments":[{"comment":"If the authors intend the 'unpredictability' to be mediated by the mean travel times themselves, they should demonstrate that a deterministic cycling schedule with identical per-driver route frequencies fails to reproduce the advantage; otherwise the claim is unsupported.","section":"Section 2, Eq. (2)"},{"comment":"Additionally, the normalized tau_avg uses t_SO_avg as the benchmark, which may mechanically favor SO algorithms; a different normalization (e.g., free-flow time) would provide a more neutral test.","section":"Experiment 3, Obj_mu"},{"comment":"This is not a request for perfection, but the discrepancy between the abstract and the body is load-bearing for the paper's central claim.","section":"Abstract and Experiment 2"},{"comment":"Without this, the 'strictly superior' claim in the Introduction is not supported by the reported evidence.","section":"Table 1 / Section 3, Experiment 2"},{"comment":"This is not a request to solve the learning problem, but the paper should not imply that the mechanism is robust to realistic information constraints.","section":"Section 3, DiscFknown=True"}],"minor_comments":[{"comment":"There are several typos and awkward phrases: 'decreseas' (p.3), 'unrelisti' (p.5), 'dependece' (p.13), 'advanteges' (p.12), 'paper paper' (p.13). Please proofread carefully.","section":"Throughout"},{"comment":"The row 'Your Algorithm ? ?' is unprofessional in a submitted manuscript; it appears to be a template placeholder. Please remove it and include a proper description of the proposed benchmark entry.","section":"Table 2"},{"comment":"The bottom panel of Figure 4 is difficult to read because the subplot structure is not clearly labeled. Please enlarge the panels and add clear axis titles for all subplots, including the one-fleet cases.","section":"Figure 4"},{"comment":"The statement 'for further experiments we choose to set the logit parameter to 0.2' is presented as a modeling decision, but the rationale is not fully quantified. Please report the actual travel-time variance for different P values so the reader can assess the trade-off.","section":"Section 3, Experiment 1"},{"comment":"Some references lack full bibliographic details (e.g., de Almeida Correia et al. 2019 is missing volume/page if available). Please check journal style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a promising benchmark framework and a clear policy idea, but the central mechanism claimed in the abstract is not actually represented in the model, and the experimental evidence is not yet at the level needed for the strong claims. The lack of code is also concerning for reproducibility, though not a formal requirement. I would encourage the authors to either (a) add a variance-sensitive HDV model or explicitly reframe the claim away from 'unpredictable travel times', and (b) provide multi-seed statistics and sensitivity analyses. With those changes, the paper could become a solid contribution. The Experiment 3 finding should be presented as a design result rather than a discovery."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a simulation benchmark for competing CAV fleets, and it shows that a randomized routing heuristic (RFlex/RFlexV) can outperform SO/UE routing for a single fleet. The SUMO appendix is a nice check. But the headline claim—that randomization wins by making travel times unpredictable for HDV drivers—is not actually represented in the model. I think the stress-test note is right.\n\nWhat is new: the dynamic two-fleet setup, the RFlex/RFlexV heuristics, and the Obj_mu objective. The authors are honest about mixed two-fleet results, and the appendix microsimulation is real effort. This is a legitimate extension of their earlier static result, not a breakthrough.\n\nWhere it is soft:\n- The HDV utility (Eq. 2) is a logit-weighted average of past mean route travel times. No variance, no day-specific information, no risk term. So a deterministic turn-taking schedule with identical per-driver frequencies would produce the same expected credibility and the same mode choices. The advantage of RFlex is therefore likely driven by selective assignment of high-gamma drivers to the faster route and by the kappa=0.5 offer, not by \"unpredictability\" per se. That weakens the abstract and the policy conclusion.\n- Experiment 2 says two-fleet results are mixed, but the abstract claims a general result. The one-fleet and two-fleet claims need to be separated.\n- No error bars, no code/data, and parameters (beta, kappa, sigma, alpha) were selected after inspecting results. The SUMO appendix has error bars for travel time but not for the modal split comparison.\n- Experiment 3 is circular: Obj_mu = (1-mu)*share + mu*tau_avg mechanically rewards lower average travel time, so finding SO- competitive at mu~0.5 is close to a property of the objective. Useful as a design illustration, but not evidence.\n\nOverall: the benchmark is usable and the one-fleet result is suggestive. The central mechanism is unproven, and the paper overstates it. A serious referee should see it because the framework could be valuable if the authors ship code, add repeated-run statistics, and reframe claims. If I were the editor, I would send it to review with a request for major revision.","headline":"A useful but overclaimed CAV routing benchmark: the one-fleet randomized routing result is suggestive, yet the headline mechanism (unpredictable travel times for HDVs) is absent from the model.","tokens_in":9371,"tokens_out":3181,"would_cite":false,"duration_ms":31825,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A80","90B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Randomized routing can win CAV fleets more customers than system-optimal routing, a simulation study claims, and mixing mean travel time into fleet payoffs can steer competition back toward social welfare.","keywords":["autonomous driving","market design","randomized routing","market share","travel time variability","day-to-day dynamics","fleet competition","traffic simulation"],"falsifier":"Run the same day-to-day benchmark but give fleet operators only observable choice histories—not the drivers' true discount factors—and require them to estimate γ per driver. If randomized routing then fails to outperform SO/UE in market share, the central claim is falsified. Alternatively, re-run with driver attitudes drawn from a low-variance distribution (e.g., all γ near 0.7); the paper's own logic predicts the randomized advantage should largely disappear.","tokens_in":8511,"feed_emoji":"🚗","tokens_out":3162,"duration_ms":26244,"temperature":0.7,"pith_summary":"This paper asks what routing algorithms competing CAV fleet operators should use when they are paid by market share, and whether that market can be steered toward city goals. It models day-to-day choices of a fixed population of drivers who can drive themselves (HDVs) or join one of two fleets, with travel times set by congestion. The paper claims that randomized routing—deliberately making travel times unpredictable for independent drivers—lets a fleet win more customers than routing proportional to system optimum or user equilibrium, provided driver attitudes toward fleets vary widely. It further claims that adding the system's average travel time to the fleets' payoff, with enough weight, restores the advantage of socially optimal routing while keeping competition alive.","feed_headline":"Randomized routing wins fleets more market share","feed_subtitle":"Model shows unpredictable travel times lure drivers; adding mean travel time to payoffs restores balance.","key_machinery":"The central objects are the randomized routing heuristics RFlexV and RFlex and the utility/credibility update of drivers. RFlexV sorts fleet members by discount factor γ and routes the highest-γ drivers to a randomly chosen route at a level that keeps them 'happy' (their disutility under fleet routing stays below HDV's). RFlex computes each driver's minimal share s_i of days on the faster route needed to keep that driver, then randomizes with target shares s_i/σ. The objective Obj_μ=(1−μ)share + μτ_avg is the device that penalizes randomization as μ grows.","core_discovery":"In a benchmark of one OD pair with two parallel routes and 200 drivers, the authors simulate two fleet operators maximizing market share. They show that the randomized heuristics RFlexV and RFlex—which keep existing customers 'happy' by routing high-discount-factor drivers on the faster route a calibrated share of days while randomizing the rest—capture more market share than proportional SO/UE routing, especially when the other fleet sticks to SO. The randomization strips human drivers of the information that one route will be faster on a given day, making independent driving less attractive. However, when the fleet objective is augmented with mean systemwide travel time (Obj_μ = (1−μ) shar","pith_inferences":["The advantage of randomization likely hinges on the assumption that fleet operators know each driver's discount factor exactly; if attitudes must be learned from observed choices, the edge may shrink or vanish, and the market-design conclusion would need re-testing.","The same logic could apply beyond CAV markets: any platform that can commit to deliberately noisy service (e.g., unpredictable pricing or delivery times) may extract market share from competitors who optimize average outcomes—suggesting regulators should watch for 'strategic opacity' as a competitive weapon.","The Obj_μ idea could be tested as a regulatory contract: city pays operators partly for low mean travel time, which the simulations suggest tames randomization without killing competition; a natural next experiment is a market with more than two routes where the SO/UE distinction matters more.","Because the paper uses two parallel identical routes, the claim that randomization is 'more efficient' is demonstrated for that topology; on asymmetric or network topologies, the effect of randomization on HDV information could differ."],"forward_implications":["If randomized routing is as effective as simulated, fleet operators maximizing market share will adopt it, and an operator facing a randomized rival cannot stay competitive with SO routing.","Market share alone as a remuneration rule invites travel-time volatility; adding average travel time to the objective, at weight roughly 0.5, makes randomization unprofitable and lets SO routing compete.","The benchmark provides a reproducible setup (BPR delay functions, logit HDV route choice, credibility update) for comparing future routing algorithms.","SUMO-based microsimulations with 150 vehicles also show the randomized algorithm beating SO, suggesting the qualitative result may extend beyond the abstract BPR setting."],"fun_headline_variants":["Randomized CAV routing captures market share","Unpredictable CAV routing beats system-optimum for fleets","Fleets gain from random routing; mean time restores balance","Market share drives CAV fleets to randomize routing","Random routing wins fleets, but system needs mean time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fleets are assumed to know each driver's discount factor toward each fleet exactly, rather than having to infer it from observed behavior; the paper's headline advantage of randomized routing is demonstrated under that knowledge, and if operators must learn attitudes from choices, that advantage is not supported.","fun_headline_variants_meta":{"raw":{"variants":["Randomized CAV routing captures market share","Unpredictable CAV routing beats system-optimum for fleets","Fleets gain from random routing; mean time restores balance","Market share drives CAV fleets to randomize routing","Random routing wins fleets, but system needs mean time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2715,"prompt_tokens":687,"completion_tokens":2028,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":1945}},"tokens_in":431,"tokens_out":2028,"duration_ms":14231,"temperature":1.0,"reasoning_tokens":1945,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:50:37.506986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same day-to-day benchmark but give fleet operators only observable choice histories—not the drivers' true discount factors—and require them to estimate γ per driver. If randomized routing then fails to outperform SO/UE in market share, the central claim is falsified. Alternatively, re-run with driver attitudes drawn from a low-variance distribution (e.g., all γ near 0.7); the paper's own logic predicts the randomized advantage should largely disappear.","supporting_citations":[],"review_version":1}