{"id":"252e80ba-1fca-4f3f-9c89-bc4a766b46e5","arxiv_id":"2608.09622","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A Bayesian MCTS planner that adapts stress decisions online improves simulated semiconductor reliability characterization yield from 20% to 54% under competing BTI, EM, and TDDB failure mechanisms.","lead":"An adaptive test planner that combines Monte Carlo tree search with Bayesian filtering chooses stress conditions step by step for semiconductor reliability tests. In simulation it raises the success rate of fully characterizing a chip from 20% to 54% while keeping competing failure mechanisms below catastrophe limits.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline improvement may be an in-sample artifact: CY is computed on the search's own rollouts on a single device, so the last-window 54% is not an unbiased estimate of a deployable policy's success rate.","rationale":"The reader's CONDITIONAL verdict is appropriate: the framework is clearly described and the simulation is reproducible in principle, but the paper's headline claim overreaches the evidence. I considered the reader's observability concern first. Section III-C and Section VIII-E explicitly state that DEM and DTDDB are treated as directly observable proxy indices and label this assumption a performance ceiling, so it is an acknowledged limitation rather than a hidden flaw. The more pressing issue is evaluation validity. Section VII-D1 and Table V compute CY on the search's own in-tree rollouts for one device instance. Because MCTS is an anytime search, the final 500 simulations of a 5,000-iteration run are not a random sample under the learned policy; they are heavily influenced by UCB exploitation of branches discovered earlier, as shown by the root entropy dropping to 0.055 and the most-visited root action taking 97% of visits. Thus the 54% figure does not measure the probability that the adaptive planner would succeed if deployed on a fresh device, or even on the same device from a cold start. The random and fixed-stress baselines are evaluated from scratch, so the comparison is not apples-to-apples. This is load-bearing because the abstract's final sentence claims that the results demonstrate significant outperformance of non-adaptive strategies for reliability qualification; a demonstration requires out-of-sample policy evaluation or at least a frozen-tree control. The paper's own Section VII-E limits the contribution to instance-conditional planning and defers population-level plans to future work, which is honest but makes the abstract's conclusion too strong. There is also an internal inconsistency in the BTI time constant: Table I gives tau ~ N(28000, 2800^2) hr while Table II lists tau = 19841 +/- 2800 hr and Table IV uses tau = 19840; this should be reconciled in revision because it affects the calibrated mechanism balance, but it is secondary to the evaluation issue. A concrete deployment-style ablation would resolve the main concern without requiring new hardware experiments.","tokens_in":19244,"tokens_out":7828,"duration_ms":76040,"concrete_test":"Run a deployment-style ablation on the same device and code: (1) use iterations 1-4,500 to build the tree; (2) freeze the tree and perform 500 rollouts using the learned UCB policy with no expansion or backpropagation, computing CY; (3) compare with the reported last-500 CY (54%) and with the baselines (16-22%). If the frozen-policy CY is close to 54%, the improvement reflects a reusable decision rule; if it drops toward 20%, the headline gain is an artifact of in-search exploitation. Additionally, repeat the comparison on 3-5 fresh device instances sampled from Table I to check instance dependence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that the central comparison is in-sample. In Section VII-D1, CY is measured on the MCTS search's own rollouts: Table V compares \"MCTS last 500\" (the final 500 simulations of a single 5,000-iteration run) with random and fixed-stress baselines, all on the same device instance. By iteration 4,500 the tree has converged: root entropy is 0.055 and the most-visited root action receives 97% of visits (Section VIII-D), so those 500 rollouts are mostly repeated exploitation of discovered high-return branches, not independent draws from a deployed adaptive policy. The baselines, by contrast, are evaluated from scratch with no search budget. The 20% to 54% improvement therefore conflates \"the search found a good path for this one device\" with \"the planner outperforms non-adaptive strategies for reliability qualification.\" This would remain a problem even under perfect observability. The reader's observability concern is real, but the paper explicitly labels the perfect-observability version a performance ceiling in Section VIII-E; the in-sample metric is a more immediate threat to the stated claim, because it affects the validity of the central comparison itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates reliability qualification of advanced semiconductor devices as a partially observable sequential decision problem in which a planner chooses stress conditions (voltage, current density, temperature, duration) to maximize the chance of characterizing BTI degradation before EM or TDDB damage becomes catastrophic. The proposed framework combines Monte Carlo tree search with a seed-action simulator (MCTS-SA) and an extended Kalman filter for online Bayesian inference over latent BTI parameters. The objective is a multi-component reward with terminal success/catastrophe terms, safety barriers, and shaping terms. The main empirical claim is that, across a single 5,000-iteration planning run on one simulated device, characterization yield rises from about 20% in the first 500 iterations to about 54% in the final 500, with 39.2% cumulative yield, and that this outperforms random and fixed-stress baselines. The paper also reports ablation studies, robustness sweeps, and an explicit limitation section.","tokens_in":19599,"tokens_out":7206,"duration_ms":69605,"significance":"The problem is relevant and the proposed formulation is a plausible extension of tree-search planning to reliability qualification. A strength of the manuscript is its candor: the direct-observability assumption is labeled a performance ceiling in Section VIII-E, the simulation-only nature of validation is stated, and the reward weights are acknowledged as calibration choices. If the reported improvement were reproduced under a properly held-out evaluation protocol and across multiple device instances, the work would be a useful contribution to adaptive test planning. However, the current evidence is in-sample, single-instance, and conditioned on ideal observability, so the significance claimed in the abstract is not yet established.","major_comments":[{"comment":"The headline comparison is in-sample. Characterization yield is measured on the MCTS search's own rollouts, and by the final 500-iteration window the tree has essentially converged: root entropy is 0.055 and the most-visited root action receives 97% of visits (Section VIII-D). The final-window CY is therefore dominated by repeated exploitation of a few discovered high-return branches for the same device instance, not by independent draws from a deployable adaptive policy. The random and fixed-stress baselines, in contrast, are evaluated from scratch with no search budget. The comparison conflates 'the search found a good path for this one device' with 'the planner outperforms non-adaptive strategies.' To support the central claim, the authors should freeze a policy from the search (e.g., the best root action or a plan derived from the converged tree) and evaluate it on held-out rollouts, ideally on fresh device instances.","section":"§VII-D1, Table V, Figure 1"},{"comment":"All quantitative results are obtained on a single device instance: Table V explicitly states 'same device instance,' and Algorithm 1 resets the environment with the same master seed each iteration. The abstract's claim that the planner 'significantly outperform[s] non-adaptive strategies for reliability qualification' is population-level in wording, but the evidence is instance-conditional. Section VII-E acknowledges the distinction and defers the population-level formulation (29) to future work, but the main conclusion is not correspondingly qualified. Reporting results over multiple sampled device instances, or at least per-instance box plots, is necessary before the stated conclusion can be drawn.","section":"§VII-D1, Table V; §VII-E"},{"comment":"The planner's safety mechanism relies on exact knowledge of D_EM and D_TDDB: the barrier penalty (Eq. 19), the damage-rate penalty, and catastrophe termination all require these values as inputs. Section VIII-E correctly states that in physical devices these indices are latent and must be estimated, and that the perfect-observability version is a performance ceiling. This is an honest limitation, but it is also load-bearing: the reported 54% final-window yield is obtained under an idealization, and the paper does not quantify how the result degrades when damage is observed through noisy indirect signals. Since the stated contribution is to reliability qualification practice, the main quantitative comparison should be complemented by a sensitivity analysis under noisy or estimated damage, or the results should be presented solely as an upper-bound benchmark.","section":"§III-C, §VIII-E, Eq. (19)"},{"comment":"The paper claims, following Ng et al., that the multi-component objective provides intermediate feedback 'without altering which test sequences are ultimately optimal.' Ng et al.'s policy-invariance theorem applies to potential-based shaping rewards of the form F = γΦ(s') − Φ(s); the additive terms r_prog, r_soft, r_prox, r_ΔD, r_U, and r_stall in Eqs. (17)–(22) are not shown to be representable as potential differences, and they accumulate along the trajectory, so they can change the total return ordering of sequences. Since the 'best successful sequence' is selected by maximizing shaped return in Eq. (27), the claim that this sequence is optimal under the terminal objective is not justified. The authors should either prove the potential-based property for their shaping terms or revise the claim to state that the components are heuristic shaping terms that trade off multiple objectives.","section":"§V, Eqs. (16)–(23), Eq. (27)"}],"minor_comments":[{"comment":"The first-500 and final-500 yields are reported inconsistently: the abstract and Figure 1 say 20.4% and 54.2%, while Table V says 19.4% and 54.0%. Please harmonize the numbers.","section":"Abstract, §VII-D1, Table V"},{"comment":"Root entropy diagnostics are inconsistent: Section VII-D1 says entropy decreases from 0.77 at iteration 100 to 0.06 by iteration 500, while Section VIII-D says it decreases from 1.0 at iteration 10 to 0.055 at iteration 5,000. Please unify the description of the convergence diagnostics.","section":"§VII-D1 vs. §VIII-D"},{"comment":"In Eq. (6), the symbol D is used for the action set of durations Δt, but D is also used throughout for damage indices D_EM and D_TDDB. Use Δt or another symbol for the action set to avoid ambiguity.","section":"§III-D, Eq. (6)"},{"comment":"The note that σ_T = σ_A = sqrt(ln(1 + 0.082)) is unclear as printed; it should read sqrt(ln(1 + 0.08²)) or the intended value should be stated explicitly.","section":"Table I"},{"comment":"The ablation statement that removing the safety barrier 'increases catastrophe rate to 78%' needs an explicit baseline comparison: 78% is close to the first-500-window rate (79.8%) and to the fixed-stress baseline (78%), and far above the final-window rate (46%). Clarify the comparison point for the ablation.","section":"§VII-D10"},{"comment":"The x-axis labels list '500' multiple times (0 500 500 1000 ...), which obscures the window boundaries. Please simplify the axis to show window indices or endpoint values.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and clearly written, but the core empirical claim rests on a single in-sample, single-instance evaluation with ideal observability. The issues are fixable: a held-out policy evaluation, multiple device instances, and a noisy-observability sensitivity analysis would directly address the main concerns. The reward-shaping claim in Section V should also be corrected or verified. The topic is within the scope of IEEE Transactions on Reliability, though the contribution is primarily a planning methodology with a simplified simulator; the authors should ensure the claims match the evidence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuine first application of MCTS-SA with EKF belief tracking to multi-mechanism reliability qualification, and the experimental write-up is unusually thorough. But the headline 20% to 54% yield improvement is not a fair comparison to the baselines: it is measured on the search's own rollouts after the tree has converged on a single device instance, so it conflates 'the search found a good path for this one device' with 'the planner outperforms static strategies.'\n\nWhat is actually new: the constrained POMDP formulation for reliability qualification, the seed-action deterministic replay that makes MCTS work with a stochastic physics simulator, and the safety-aware reward shaping with progressive widening. The paper does a solid job of calibrating the simulator so BTI, EM, and TDDB compete on comparable timescales, and it backs the central mechanism with ablations: removing the safety barrier pushes catastrophe rate to 78%, removing early shaping cuts CY to 31%, and the per-epoch cost reduces test length without hurting yield. The robustness sweeps, grid-sensitivity analysis, and anytime-behavior discussion are all real value.\n\nThe soft spots, in order of severity. First, the in-sample evaluation. Table V compares MCTS last-500 against random and fixed-stress baselines, but those last 500 rollouts come from a tree whose root has 97% of visits on one action; they are exploitation of discovered branches, not independent draws from a deployable policy. The baselines are evaluated from scratch with no search budget. So the 54% versus 16%/22% comparison is not apples-to-apples. This is fixable: freeze the final policy (e.g., greedy action selection from the tree) and evaluate it on held-out episodes, ideally across a sample of device instances from the population. Second, the headline rests on a single device instance; the parameter-shift sweeps in Table IX are informative but are not a population evaluation. Third, the reward weights are hand-tuned on the same simulator used for evaluation; the ablations help, but there is no independent calibration set. Fourth, minor: the BTI time constant is 28,000 hr in Table I but 19,841 hr in Table II and IV; the EKF prior is set to 28,000, so the filter starts with a biased mean. Fifth, the perfect observability of DEM and DTDDB is a real practical limitation, but the paper explicitly calls it a performance ceiling in Section VIII-E, so I would treat it as context rather than a fatal flaw.\n\nWho should read it: reliability engineers who want to see a concrete sequential-planning formulation, and MCTS/RL practitioners interested in a realistic constraint-heavy application. It deserves a serious peer review: the problem is important and the presentation is coherent. But the central claim needs to be re-evaluated with a deployed policy on multiple instances before it can stand. I would send it out with a request for major revision.","headline":"A serious, well-organized first application of MCTS to adaptive reliability testing, but the headline yield gain is an in-sample artifact of a single-device search and needs a deployed-policy, multi-instance evaluation.","tokens_in":20158,"tokens_out":5591,"would_cite":true,"duration_ms":47056,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sequential Bayesian planning can synthesize damage-aware stress policies that significantly outperform non-adaptive strategies for multi-mechanism reliability qualification under competing failure modes.","keywords":["reliability qualification","adaptive test planning","Monte Carlo tree search","extended Kalman filter","competing failure modes","bias temperature instability","electromigration","time-dependent dielectric breakdown"],"falsifier":"Run the same planner on the same device population with the damage indices corrupted by realistic measurement noise or replaced by estimates from a sensor model: if the characterization-yield advantage over a fixed-stress plan shrinks to zero, or if the learned sequences no longer keep $D_{\\mathrm{EM}}$ and $D_{\\mathrm{TDDB}}$ below 1, the central claim that adaptive planning significantly outperforms non-adaptive strategies in the intended setting would be refuted.","tokens_in":18975,"feed_emoji":"⚙️","tokens_out":5983,"duration_ms":48840,"temperature":0.7,"pith_summary":"Reliability qualification of advanced semiconductor devices currently uses fixed stress recipes that ignore per-device variability and competing failure mechanisms. This paper tries to establish that closed-loop sequential planning can do better: it frames qualification as a partially observable decision problem and solves it by combining Monte Carlo tree search with an extended Kalman filter that tracks latent degradation parameters. In a physics-based simulation of concurrent BTI, electromigration, and time-dependent dielectric breakdown, the adaptive planner's characterization yield rises from about 20% in the first 500 planning iterations to over 54% in the final 500, reaching a 39% cumulative success rate while keeping the best successful sequence well inside safety margins. The claim is that these gains come from learned, damage-aware stress policies rather than from random exploration, and that the approach outperforms non-adaptive baselines.","feed_headline":"Adaptive test planner lifts reliability yield from 20% to 54%","feed_subtitle":"Closed-loop MCTS steers accelerated stress between competing failure modes, beating static test recipes.","key_machinery":"The load-bearing mechanism is Monte Carlo tree search for seed-action simulators (MCTS-SA) coupled with extended Kalman filter (EKF) belief estimation. Each decision is a pseudorandom seed that maps deterministically to a stress action $(V,J,T,\\Delta t)$, so stochastic simulations can be replayed exactly and planning statistics stay consistent. The EKF maintains a posterior over the latent BTI parameters $(\\Delta V_{\\max},\\tau,\\beta)$ and supplies the trace of its covariance as an uncertainty signal. A shaped reward combines progress toward the characterization threshold, an early-progress bonus, a cubic safety barrier that penalizes damage above $D_{\\mathrm{thr}}=0.25$, per-epoch damage and uncertainty penalties, and a terminal reward that weights successful characterization at 20,000 against catastrophic failure at 2,000. The interaction of progressive widening with the seed-action space keeps the tree size governed by visit counts rather than by the raw action-grid cardinality, which is what makes finer stress grids nearly free.","core_discovery":"The paper's central claim is that multi-mechanism reliability qualification can be treated as a constrained sequential decision problem and solved with tree-search planning, even though degradation is irreversible and the latent device parameters are unknown. The planner chooses stress conditions from a discrete grid, observes noisy threshold-voltage shifts, and updates a Gaussian belief over BTI parameters with an EKF; EM and TDDB damage indices act as directly observable safety constraints. Across 5,000 planning iterations the characterization yield improves monotonically by window, from 20.4% to 54.0%, with cumulative yield 39.2% and a final-window mean return that turns positive. The best successful sequence terminates at epoch 119 with $\\Delta V_t \\ge 0.09$ V, posterior uncertainty $U=0.044$, and damage fractions $D_{\\mathrm{EM}}=0.564$ and $D_{\\mathrm{TDDB}}=0.537$, which the paper reads as evidence that the search learns to thread the narrow safe corridor between characterization and catastrophe.","pith_inferences":["Beyond the paper: a realistic deployment would need to estimate the damage indices from sensor signals rather than assume perfect observability, likely eroding some of the demonstrated yield gain until the estimator is added.","Beyond the paper: the same formulation should transfer to battery aging or structural fatigue, where multiple competing wear-out mechanisms constrain accelerated testing.","Beyond the paper: a population-level policy that marginalizes over device parameters, plus distillation of the tree search into a fast approximator, would convert instance-optimal sequences into reusable real-time test recipes."],"forward_implications":["Static qualification recipes derived from population averages can be replaced by per-unit adaptive policies that condition each stress decision on the measurements already taken.","The learned strategy of moderating stress as damage approaches the barrier, and preferring temperature over voltage when damage is high, transfers as a design principle for test protocols.","Because the planner is anytime, a usable test plan is available well before full convergence, letting the planning budget be matched to available test time.","Finer discretization of the stress grid improves achievable yield without combinatorial growth in search-tree size, because progressive widening bounds children by visit count.","Ablation shows the safety barrier and early-progress shaping are load-bearing: removing the barrier raises the catastrophe rate to 78% and removing early shaping drops yield to 31%."],"supporting_citations":[{"why":"Supplies the physics-based temporal simulator used to model concurrent BTI, EM, and TDDB degradation with per-device variability.","marker":"[15]"},{"why":"Introduces the adaptive stress testing and seed-action formulation that make deterministic replay in MCTS possible.","marker":"[11]"},{"why":"Surveys black-box safety validation algorithms and seed-action MCTS, grounding the planning method and progressive widening.","marker":"[12]"},{"why":"Provides the stretched-exponential BTI compact model and the basis for using absolute threshold-voltage shift as the characterization metric.","marker":"[4]"},{"why":"Supplies physics-based NBTI models used for the BTI degradation dynamics and parameter identifiability reasoning.","marker":"[5]"},{"why":"Reports 14nm FinFET reliability data motivating simultaneous consideration of BTI, EM, and TDDB.","marker":"[1]"},{"why":"Reports 10nm FinFET reliability competition used to calibrate mechanism timescales.","marker":"[2]"},{"why":"Formalizes the POMDP belief-state foundation for acting under partial observability.","marker":"[10]"},{"why":"Supplies the extended Kalman filter machinery for online Bayesian belief updates.","marker":"[13]"},{"why":"Provides the MCTS survey underpinning selection, progressive widening, and anytime behavior.","marker":"[14]"}],"fun_headline_variants":["Tree-search planner adapts stress tests, boosting yield to 54%","Bayesian MCTS steers stress tests to 54% characterization yield","Closed-loop planner balances failure modes, lifting yield to 54%","Adaptive tests beat static recipes: yield hits 54% from 20%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The planner assumes it sees the true EM and TDDB damage indices directly at every decision step, although in real devices those quantities are latent and can only be estimated from indirect electrical signals; the paper calls this perfect-observability version a performance ceiling.","fun_headline_variants_meta":{"raw":{"variants":["Tree-search planner adapts stress tests, boosting yield to 54%","Bayesian MCTS steers stress tests to 54% characterization yield","Closed-loop planner balances failure modes, lifting yield to 54%","Adaptive tests beat static recipes: yield hits 54% from 20%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":2036,"prompt_tokens":1036,"completion_tokens":1000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":920}},"tokens_in":652,"tokens_out":1000,"duration_ms":7198,"temperature":1.0,"reasoning_tokens":920,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:55:56.406074+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same planner on the same device population with the damage indices corrupted by realistic measurement noise or replaced by estimates from a sensor model: if the characterization-yield advantage over a fixed-stress plan shrinks to zero, or if the learned sequences no longer keep $D_{\\mathrm{EM}}$ and $D_{\\mathrm{TDDB}}$ below 1, the central claim that adaptive planning significantly outperforms non-adaptive strategies in the intended setting would be refuted.","supporting_citations":[{"cited_title":"Gerabaldi: A temporal simulator for probabilistic ic degradation and failure processes,","cited_arxiv_id":null,"evidence_quote":"Supplies the physics-based temporal simulator used to model concurrent BTI, EM, and TDDB degradation with per-device variability."},{"cited_title":"Adaptive Stress Testing: Finding Likely Failure Events with Reinforcement Learning","cited_arxiv_id":"1811.02188","evidence_quote":"Introduces the adaptive stress testing and seed-action formulation that make deterministic replay in MCTS possible."},{"cited_title":"A survey of algorithms for black-box safety validation of cyber-physical systems,","cited_arxiv_id":null,"evidence_quote":"Surveys black-box safety validation algorithms and seed-action MCTS, grounding the planning method and progressive widening."},{"cited_title":"The paradigm shift in understanding the bias temperature instability: From reaction-diffusion to switching oxide traps,","cited_arxiv_id":null,"evidence_quote":"Provides the stretched-exponential BTI compact model and the basis for using absolute threshold-voltage shift as the characterization metric."},{"cited_title":"A comparative study of different physics-based NBTI models,","cited_arxiv_id":null,"evidence_quote":"Supplies physics-based NBTI models used for the BTI degradation dynamics and parameter identifiability reasoning."},{"cited_title":"Systematical study of 14nm FinFET reliability: From device level stress to product HTOL,","cited_arxiv_id":null,"evidence_quote":"Reports 14nm FinFET reliability data motivating simultaneous consideration of BTI, EM, and TDDB."},{"cited_title":"Reli- ability studies of a 10nm high-performance and low-power CMOS tech- nology featuring 3rd generation FinFET and 5th generation HK/MG,","cited_arxiv_id":null,"evidence_quote":"Reports 10nm FinFET reliability competition used to calibrate mechanism timescales."}],"review_version":1}