{"id":"ad09f80e-1a7e-4aee-85e9-a1305f233c47","arxiv_id":"2607.11075","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Joint-TS samples MCS success probabilities from a multivariate ordered Beta prior via Gibbs sampling, preserving monotonicity and delivering robust link-adaptation throughput across CQI and Doppler settings.","lead":"Joint-Thompson Sampling picks wireless modulation rates while forcing success probabilities to stay ordered, using a multivariate ordered Beta prior and Gibbs sampling. It keeps throughput competitive and stable when CQI is missing or Doppler is high, where several existing bandit methods degrade.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Fixed N=1000 Gibbs sweeps lack evidence of adequate mixing for the K=29 MCS case that underpins all reported throughput results.","rationale":"The reader correctly isolates the finite-N Gibbs approximation as the softest link supporting the strongest claim. The mathematics of the MOB and the Gibbs conditionals (Lemma 1) are sound, the critique of CoTS (Remark 2) is accurate, and the reported tables do show Joint-TS avoiding the sharp drops of UTS (no-CQI) and LTS (high Doppler). Those empirical patterns survive even if the sampler is only approximate; they simply cannot be confidently credited to the joint ordered prior until mixing for K=29 is verified. Because the paper already makes no regret claims, ships no code, and evaluates only single-path channels, the existing CONDITIONAL verdict already accounts for the remaining experimental limitations. Confirming or refuting the sampler quality would either strengthen the attribution of robustness to MOB or reveal that the gains come from something else (smoothing, initialization, etc.), but it does not move the overall recommendation.","tokens_in":9804,"tokens_out":602,"duration_ms":28363,"concrete_test":"After 200 TTIs of a representative no-CQI Doppler run (Fd=20 Hz, avg SINR 10 dB), freeze the current (α,β). Draw 500 independent sample vectors with N=1000 and another 500 with N=20 000. Compute the fraction of draws for which arg max_i s_i Θ_i differs between the two N values. If that fraction exceeds 5 %, re-run the full Table II experiments with N=20 000; any throughput shift >0.05 bps/Hz shows that the published numbers depend on an under-mixed sampler.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Joint-TS is robust because it samples from the true MOB posterior rests on Algorithm 2 with a hard-coded N=1000. Lemma 2 only guarantees total-variation convergence as N\to∞; the sole empirical check is a visual 2-D histogram for K=2 (Fig. 1c). For the K=29 5G NR MCS set actually used in every simulation (static and Doppler, with/without CQI), no mixing-time bound, effective-sample-size diagnostic, or decision-stability check is supplied. Because each online decision re-initializes from ordered uniforms and then takes only 1000 sequential neighbor updates, the generated vectors may still be far from the target MOB when the Beta parameters become large or unbalanced. If those vectors systematically deviate from the ordered posterior, the claimed advantage of the joint prior over independent TS or CoTS is not actually realized, and the observed robustness cannot be attributed to the method as described.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper formulates MCS selection for link adaptation as a multi-armed bandit and proposes Joint-TS, which places a multivariate ordered Beta (MOB) prior on the vector of success probabilities so that samples always obey the natural monotonicity θ1 ≥ ··· ≥ θK. Conjugacy of the MOB under Bernoulli observations is shown, and sampling is performed by Gibbs iteration over restricted-Beta conditionals (Algorithm 2, Lemma 1). Empirical comparisons against OLLA, MTS, UTS and LTS are reported for static SINR and for three Doppler values, both with perfect CQI and without CQI, using the 29 MCS indices of 5G NR. The central claim is that Joint-TS yields competitive throughput and is more robust across these regimes than the listed baselines.","tokens_in":10039,"tokens_out":1149,"duration_ms":18575,"significance":"If the sampler is faithful and the empirical gains hold under more realistic channel models, the work supplies a clean, table-free Bayesian alternative to OLLA and to independent-Beta Thompson sampling that correctly exploits monotonicity. The conjugacy argument and the explicit contrast with the non-joint sequential sampler of CoTS [18] are technically sound contributions. The multi-scenario simulation suite (static, Doppler, CQI/no-CQI) is a useful practical stress test for the 5G/6G community. The absence of regret bounds is acceptable for an applications-oriented paper, but the lack of any mixing diagnostic for the K=29 case that underpins every reported number currently limits the strength of the attribution of robustness to the joint prior.","major_comments":[{"comment":"Sec. IV-C, Lemma 2 and Algorithm 2: Lemma 2 only guarantees total-variation convergence of the Gibbs chain as N\to∞. The sole empirical check is a visual 2-D histogram for K=2 (Fig. 1c). All throughput results, however, use K=29 MCS indices with a hard-coded N=1000 restarts from ordered uniforms at every TTI. No effective-sample-size, autocorrelation, or decision-stability diagnostic is supplied for the parameter regimes that arise after many updates (large or unbalanced α,β). Without such evidence the claim that Joint-TS realises the true MOB posterior—and that this is the source of its robustness—remains unverified for the experimental setting that matters.","section":null},{"comment":"Sec. V and Table II: The robustness narrative rests on averages over only 20 independent trials and a single-path channel model. For the high-Doppler (111 Hz) no-CQI cell the absolute throughputs of UTS, LTS and Joint-TS are close (1.34 / 1.29 / 1.32 bps/Hz); it is unclear whether the ranking is statistically stable. A modest increase in trial count, multi-path models, and confidence intervals would be needed before the “consistent performance in all scenarios” claim can be regarded as established.","section":null},{"comment":"Sec. V (CQI-less paragraph): The exponential forgetting window w=50 is fixed by hand and the text defers adaptive choice of w to future work. Because the same w is used for the smoothed MTS/UTS/Joint-TS variants, any sensitivity of the ranking to w is unexplored; a short ablation would strengthen the comparison.","section":null}],"minor_comments":[{"comment":"Fig. 2 captions and Table II: units and averaging horizon (1000 TTIs) should be stated uniformly in both places; the Oracle curves in panels (b)–(c) are not defined in the caption.","section":null},{"comment":"Remark 1 and Algorithm 2: the special case of identical (α,β) that reduces to sorting i.i.d. Betas is useful; a one-line note that the general Gibbs still applies when some α i,βi become large would help implementers.","section":null},{"comment":"Typographical: “pyitpp” / “py-itpp” inconsistency; missing space in “ai .iith.ac.in”; “BLER(i,c)” sometimes written without spaces.","section":null},{"comment":"Related-work paragraph: a brief quantitative comparison of wall-clock cost of N=1000 Gibbs sweeps versus independent Beta sampling would clarify real-time feasibility at the MAC layer.","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical core (MOB conjugacy + correct Gibbs) is solid and the distinction from CoTS is fair. The main risk is that the empirical story is currently under-powered and the sampler quality for K=29 is unexamined; both are fixable within a revision cycle. Scope is appropriate for a communications/ML systems venue; novelty relative to [18] is incremental but the Doppler/CQI focus and the proper joint sampler are genuine increments."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real addition here is a Gibbs sampler that actually targets the multivariate ordered Beta joint, plus a clean empirical showing that the resulting Joint-TS stays competitive when CQI disappears or Doppler is high. They correctly flag that CoTS’s sequential inverse-transform does not sample the true MOB (Remark 2 is the sharpest paragraph in the paper), and they give the right conditional restricted-Beta updates (Lemma 1) and conjugacy. That is new relative to Gupta et al. and to the independent-Beta TS variants they compare against.\n\nWhat they do well is the systems evaluation. Using the pyitpp library they run the full 29-MCS 5G NR table under static SINR and three Doppler values, both with perfect CQI and without. Table II and Fig. 2 make the robustness claim concrete: UTS collapses without CQI, LTS degrades at high Doppler, Joint-TS does not. The algorithm itself is simple (Algorithm 1) and the prior literature is cited fairly.\n\nThe soft spot is exactly the one the stress-test flags. Lemma 2 only gives asymptotic total-variation convergence; the only diagnostic is a K=2 histogram. Every throughput number rests on N=1000 Gibbs sweeps for K=29 with no ESS, no decision-stability check, and no mixing-time argument. When the Beta parameters become large or unbalanced the samples may still be far from the target posterior, so the claimed advantage of the joint prior is not fully secured. Minor additional gaps: no regret bound, no code, no error bars, single-path channels only, and a hand-tuned smoothing window w. None of these sink the paper; they just keep it conditional.\n\nThis is for people who already work on MAB rate adaptation or OLLA replacements. A serious referee should see it. I would cite the sampling correction and the CQI-less Doppler numbers if I were writing in the same area. Send it to peer review.","headline":"Correct Gibbs sampler for ordered Beta priors in MCS selection; useful robustness data, but N=1000 mixing for K=29 is unproven.","tokens_in":10664,"tokens_out":493,"would_cite":true,"duration_ms":4389,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Joint Thompson sampling that respects ordered MCS success rates delivers competitive link throughput in every tested regime.","keywords":["Thompson Sampling","multi-armed bandits","link adaptation","MCS selection","ordered Beta distribution","Gibbs sampling","5G NR","Wi-Fi"],"falsifier":"Re-run the same 5G-NR Doppler simulations with N reduced to a few tens of Gibbs iterations (or with an exact MOB sampler) and check whether the reported throughput gaps versus UTS/LTS reverse or disappear.","tokens_in":10703,"feed_emoji":"📡","tokens_out":545,"duration_ms":4818,"temperature":0.7,"pith_summary":"Wireless transmitters must pick a modulation and coding scheme (MCS) on every packet so that throughput is high while errors stay rare. The problem is usually cast as a multi-armed bandit in which each MCS is an arm whose unknown success probability is learned from ACK/NACK feedback. Classical Thompson sampling treats those probabilities as independent Beta random variables and therefore can propose impossible orderings (a higher-rate MCS succeeding more often than a lower-rate one). This paper replaces the independent priors with a single multivariate ordered Beta distribution that forces the sampled success probabilities to remain monotonically decreasing. The resulting Joint-TS algorithm is shown, across static channels and across Doppler scenarios with and without CQI feedback, to match or exceed existing bandit link-adaptation methods while never collapsing in the regimes where those methods fail.","feed_headline":"Ordered Thompson sampling keeps link rates high in every regime","feed_subtitle":"Joint-TS never collapses when CQI vanishes or Doppler rises, matching the best prior bandits","key_machinery":"Multivariate ordered Beta (MOB) prior together with a Gibbs sampler that draws from the correct restricted-Beta conditionals; each draw automatically lies inside the ordered simplex and the posterior stays MOB after a Bernoulli observation.","core_discovery":"By sampling the vector of MCS success probabilities jointly from a multivariate ordered Beta prior (instead of independent Betas), Thompson sampling automatically respects the physical ordering of the arms; the posterior remains ordered after every ACK/NACK update, and the resulting online policy yields competitive spectral efficiency in every static and time-varying setting examined.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Joint-TS preserves MCS order for robust link adaptation","Ordered Beta prior keeps Thompson sampling consistent across regimes","Joint sampling succeeds where independent MCS arms fail","Multivariate ordered prior yields competitive rates in all settings","Joint-TS avoids collapse when CQI vanishes or Doppler rises"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a fixed number of Gibbs steps (N=1000) produces samples close enough to the true ordered posterior for every online MCS decision to stay near-optimal.","fun_headline_variants_meta":{"raw":{"variants":["Joint-TS preserves MCS order for robust link adaptation","Ordered Beta prior keeps Thompson sampling consistent across regimes","Joint sampling succeeds where independent MCS arms fail","Multivariate ordered prior yields competitive rates in all settings","Joint-TS avoids collapse when CQI vanishes or Doppler rises"]},"model":"grok-4.5","effort":"low","cost_usd":0.00379,"raw_usage":{"total_tokens":1183,"prompt_tokens":731,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":37900000,"prompt_tokens_details":{"text_tokens":731,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":394,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":731,"tokens_out":58,"duration_ms":3636,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T07:13:39.249283+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same 5G-NR Doppler simulations with N reduced to a few tens of Gibbs iterations (or with an exact MOB sampler) and check whether the reported throughput gaps versus UTS/LTS reverse or disappear.","supporting_citations":[],"review_version":1}