Pith

open record

sign in

arxiv: 2607.05805 · v1 · pith:EJKBIPEF · submitted 2026-07-07 · cs.AI · cs.LG· quant-ph

Onnes: A Physics-Grounded Multi-Agent LLM Simulator for Cryogenic Fault Diagnosis in Quantum Computing Infrastructure

Praneeth Narisetty , Uday Kumar Reddy Kattamanchi , Shiva Nagendra Babu Kore This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 reserved 2026-07-08 23:30 UTCglm-5.2pith:EJKBIPEFrecord.jsonopen to challenge →

Figure 1
Figure 1. Figure 1: End-to-end system. The twin (a) renders a telemetry window from one seed; a single [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] reproduced from arXiv: 2607.05805
classification cs.AI cs.LGquant-ph
keywords dilution refrigeratorfault diagnosisLLM agentsin-context learningdigital twinquantum computing infrastructureself-consistencylabel efficiency
0
0 comments X

The pith

Six labeled examples match 300-label ML on cryogenic fault diagnosis

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents Onnes, a physics-grounded simulator of a dilution refrigerator (the cryogenic system that houses superconducting quantum computers) paired with a multi-agent LLM operations layer for fault diagnosis. The central claim is that a zero-shot LLM agent panel, given just six curated contrastive demonstrations and self-consistency voting, matches a supervised random forest classifier (0.990 vs. 0.985 accuracy) on six physics-grounded fault classes, with no parameter updates. The paper builds a digital twin that couples a real dilution-cooling model with noise fingerprints learned from actual BlueFors refrigerator logs, and engineers three fault classes to be deliberately confusable on temperature while separable only on flow and pressure channels. Zero-shot, the agent panel shows no statistically significant difference from supervised ML on fault detection but loses on classification, with all errors concentrated on the confusable pairs. The contrastive demonstrations close exactly those error cells. An ablation attributes the lift almost entirely to the few-shot demonstrations rather than self-consistency voting, and shows that a single LLM call matches the five-role panel on accuracy, meaning the multi-agent structure earns its cost on auditability and a safety veto surface rather than classification performance. As a first sim-to-real check, an anomaly detector trained purely on real BlueFors telemetry achieves a 6.4% false-alarm rate and 100% recall on physics faults injected onto real held-out windows.

Core claim

The load-bearing discovery is that the entire classification gap between a zero-shot LLM agent and a supervised classifier on cryogenic fault diagnosis is concentrated on physically confusable fault pairs (helium leak vs. blocked impedance; wiring heat ingress vs. heat load spike), and that six contrastive demonstrations curated by an a priori thermal-degeneracy prior close that gap completely. The ablation reveals that few-shot demonstrations are the load-bearing mechanism (0.500 to 0.983 alone), self-consistency voting adds nothing on its own (0.483), and the five-role panel buys no accuracy over a single well-prompted call, earning its cost only on separable safety surfaces: an inspecting

What carries the argument

Onnes digital twin: a forward physics model of dilution refrigeration (T-squared cooling floor) with a learned real-fridge noise fingerprint from BlueFors logs, six physics-grounded fault classes (three engineered to overlap on temperature but separate on flow and pressure), and a five-role LLM agent panel (Sentinel, Diagnostician, Operator, Guardian, Supervisor) with optional contrastive few-shot demonstrations and self-consistency voting attached to the Diagnostician.

If this is right

  • For newly commissioned dilution refrigerators with no fault history, six curated demonstrations could substitute for hundreds of diagnosed fault episodes, making first-deployment fault diagnosis feasible before labeled data accumulates.
  • The multi-agent decomposition's value is auditable safety, not accuracy: a single LLM call with the same demonstrations matches the panel, suggesting production systems should default to the cheaper single-call path and escalate to the full panel only when a safety veto or audit trail is needed.
  • The confidence-gating mechanism for continuous monitoring trades false alarms for detection latency at zero additional cost, and its false-alarm rate is backend-dependent, meaning the gate should ship as a tunable knob rather than a fixed threshold.
  • The sim-to-real transfer result (6.4% false-alarm rate on real telemetry, 100% recall on injected physics faults) suggests the detection layer survives real hardware noise, though classification on real faults remains unvalidated.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 0.990 accuracy is explicitly described by the authors as an optimistic upper bound because the few-shot demonstration curation (k=6, confusable-pair weighting, tau=0.7) was fixed using knowledge of where the zero-shot panel fails on the evaluation scenarios rather than on a separate tuning split. If a blind dev/eval split were to show that the curation does not generalize to unseen fault distr
  • The twin's fault signatures are authored at high signal-to-noise ratio, which is why both the enhanced panel and supervised ML saturate near ceiling. The more informative comparison regime may be the high-sensor-noise stress test where the random forest holds 0.701 macro-F1 at 20% noise, suggesting the label-efficiency advantage of the agent may be larger in harder, non-saturated regimes.
  • The twin is not a statistically indistinguishable replica of any specific fridge: a classifier two-sample test achieves AUC=1.00 separating twin windows from real BlueFors windows. All accuracy numbers should be read as 'on the twin,' and the transferable claim is the relative comparison structure (agent vs. ML at equal label budget) rather than absolute performance.

Load-bearing premise

The six contrastive demonstrations were curated using knowledge of where the zero-shot panel fails on the evaluation scenarios rather than on a separate tuning split, so the 0.990 accuracy is best read as a configuration tuned toward this evaluation set rather than evidence of transferable capability.

What would settle it

A blind dev/eval split where the demonstration curation is fixed on a development set and evaluated on a held-out set with a different fault distribution would test whether the parity result generalizes; if the enhanced panel's accuracy drops substantially below the supervised baseline on unseen distributions, the parity claim does not hold.

Figures

Figures reproduced from arXiv: 2607.05805 by Praneeth Narisetty, Shiva Nagendra Babu Kore, Uday Kumar Reddy Kattamanchi.

Figure 2
Figure 2. Figure 2: Data flow of the twin. Physics sets the mean trajec [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quench physics, corrected. Steady-load model: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Three thermal faults overlap on MXC temperature [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Baseline zoo on identical n=200 held-out seeds (sev × 0.5), with 95% CIs on accuracy. TabPFN-2.5 is the strongest opponent; we keep the random forest as the primary head-to-head opponent because it is fully local and repro￾ducible, and report TabPFN-2.5 as the stronger foundation￾model bar [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Supervised ML under realism stress. Macro- [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Clopper–Pearson 95% confidence intervals on ac￾curacy (n=200; detection accuracy 0.965/0.995, classifica￾tion accuracy 0.685/0.985 for agent/RF). These are the raw￾accuracy proportions the exact tests run on, and differ from the detection F1 reported in [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Head-to-head. Detection shows no significant differ [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Zero-shot agent confusion matrix on the n=200 held-out scenarios (rows: true class; columns: the Supervi￾sor’s predicted class; cells: scenario counts). The two dom￾inant off-diagonal cells are the engineered confusable pairs — helium_leak→blocked_impedance (23 cases) and wiring_heat_ingress→heat_load_spike (16 cases) — which the twin was built to make ambiguous on temperature; the off-diagonal mass there… view at source ↗
Figure 11
Figure 11. Figure 11: Confusion matrices with absolute counts ( [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: From raw window to verdict for one real scenario (eval seed [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 15
Figure 15. Figure 15: Confidence gating trades precision for latency, [PITH_FULL_IMAGE:figures/full_fig_p010_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Label efficiency: multiclass accuracy (y) versus number of labeled training scenarios (x, log scale) on identical held-out seeds (severity fixed; 6 seeds/cell). Shaded bands are 95% CIs over the 6 trainings per cell; the dashed violet line is the enhanced panel’s six-demonstration accuracy (0.990) and the dotted grey line the zero-shot panel (0.685), both training-free references (flat because they use no… view at source ↗
read the original abstract

Dilution refrigerators are the enabling infrastructure of superconducting quantum computers, yet their fault diagnosis is still dominated by threshold alarms that report that something is wrong, not what. We present Onnes, a physics-grounded digital-twin simulator of a dilution refrigerator (a forward physics model with a learned real-fridge noise fingerprint) that drives a live multi-agent LLM operations layer, and use it for a controlled head-to-head between a zero-shot LLM agent panel and a supervised ML classifier on cryogenic fault diagnosis. The twin couples a real dilution-cooling floor, a noise-and-correlation fingerprint learned from real BlueFors logs, and six physics-grounded fault classes, three engineered to overlap on temperature but separate on flow and pressure. Across a 1000-turn evaluation the zero-shot panel shows no significant difference from the classifier on detection but trails on classification, its errors concentrating on the confusable faults. Curated contrastive few-shot demonstrations and self-consistency voting then raise classification accuracy from 0.685 to 0.990, matching the supervised classifier (0.985) with no parameter updates and six labeled demonstrations; an ablation attributes the gain almost entirely to the demonstrations. Run as a continuous monitor across a nine-run fault-by-seed sweep, the agent catches every developing fault within one poll interval, and a confidence gate suppresses pre-onset false alarms whose rate is backend-dependent. As a first sim-to-real check, a detector trained purely on real BlueFors telemetry posts a real-hardware false-alarm rate of 6.4% and 100% recall on physics faults injected onto real held-out windows. All numbers are drawn verbatim from released run logs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 6 minor

Summary. The paper presents Onnes, a physics-grounded digital-twin simulator for dilution refrigerators coupled to a five-role LLM agent panel for cryogenic fault diagnosis. The twin combines a T² dilution-cooling floor, a noise fingerprint learned from real BlueFors logs, and six fault classes engineered so that three (helium_leak, blocked_impedance, wiring_heat_ingress) are confusable on temperature but separable on flow and pressure. The central empirical result is a controlled head-to-head: zero-shot, the agent panel matches a supervised random forest on detection (McNemar p=0.07) but trails on classification (0.685 vs. 0.985); curated contrastive few-shot demonstrations plus self-consistency voting then raise classification to 0.990, matching the RF (0.985) with six labeled demonstrations and no parameter updates. An ablation attributes the lift almost entirely to the demonstrations. A 9-run continuous-monitoring sweep and a first sim-to-real check (6.4% false-alarm rate on real BlueFors telemetry) are also reported. All numbers are drawn from released run artifacts.

Significance. The paper ships several commendable artifacts: (1) a released, reproducible digital-twin and evaluation harness with a claim-to-artifact map (Table 9); (2) a baseline zoo including TabPFN-2.5 and an engineered physics-rule baseline, not a straw man; (3) exact paired McNemar tests and Clopper-Pearson CIs throughout; (4) a cross-backend replication (Claude vs. Gemini, Table 8); (5) an ablation that honestly reports the multi-agent structure buys no accuracy over a single call (Table 6); and (6) a falsifiable sim-to-real validation plan with stage 1 already executed. The negative results (avoiding debate/self-refinement; multi-agent structure does not improve accuracy) are grounded in 2026 literature and reported as findings rather than buried. The label-efficiency sweep (Fig. 16) quantifies the 6-vs-300 example-count gap concretely. These are substantive contributions to the methodology of evaluating LLM agents on physics-constrained diagnostic tasks.

major comments (1)
  1. §9, Table 10, and §7: The headline parity claim (0.990 vs. 0.985, Table 5) depends on few-shot hyperparameters (k=6, N=3, τ=0.7) and confusable-pair demonstration weighting that the authors acknowledge were 'fixed using knowledge of where the zero-shot panel fails on the evaluation scenarios, not on a separate tuning set' (§9). The authors frame 0.990 as 'an optimistic upper bound,' which is the right characterization, but the abstract and conclusion still lead with 'parity without training' as the headline. The concern is bounded by the ablation (Table 6: few-shot alone reaches 0.983 on a harder seed set) and by cross-backend replication (Table 8: Gemini 0.715→0.995 with the same configuration), both of which suggest the result is not fragile to the exact lever combination. However, neither test addresses the core issue: the same tuned hyperparameters were applied to both backends and a
minor comments (6)
  1. §3, Eq. (1): the cold-plate formula 'TCP = 0.1√(Q̇/Q̇₁₀₀)' is introduced without a clear definition of Q̇₁₀₀ in the immediate context; the reader must infer it from the surrounding text. A one-line definition would help.
  2. Table 3: detection F1 (0.979 vs. 0.997) is reported alongside classification accuracy, but Fig. 9 reports detection accuracy (0.965/0.995). The relationship between these two detection metrics is not immediately obvious; a footnote or cross-reference would prevent confusion.
  3. §8: the 9-run sweep is described as '3 fault classes × 3 seeds = 9 continuous 24 h runs,' but the specific fault classes and seeds are not listed in the text or a table. Adding these would improve reproducibility beyond the released JSON.
  4. Table 10: the LLM temperature and top_p are listed as 'unset (backend default)' for Claude but the self-consistency diversity temperature τ=0.7 is listed separately. It would help to clarify whether τ=0.7 overrides the backend default or is applied only during the N=3 sampling.
  5. §9: the twin-fidelity discriminator (AUC=1.00) is a striking result that deserves slightly more discussion. The authors note the fingerprint matches noise magnitudes and correlations 'to within a few percent' but a classifier still separates perfectly. A sentence on which features drive the separation (the authors mention feature importances name the channels) would strengthen the honesty of this section.
  6. Appendix A.7: the Selective Verifier prompt is released but 'off by default.' It would help to state in the main text (§7) that the verifier is not used in any headline result, to avoid confusion for readers who skim the appendix.

Simulated Author's Rebuttal

2 responses · 0 unresolved

The referee recommends minor revision and identifies one major comment (truncated in the report) concerning the headline parity claim's dependence on few-shot hyperparameters tuned with knowledge of evaluation-set failures rather than on a separate dev split. We agree this is a legitimate concern and will revise the abstract and conclusion to lead with the 'optimistic upper bound' framing rather than 'parity without training,' while preserving the substantive results the referee commends.

read point-by-point responses

  1. Referee: §9, Table 10, and §7: The headline parity claim (0.990 vs. 0.985, Table 5) depends on few-shot hyperparameters (k=6, N=3, τ=0.7) and confusable-pair demonstration weighting that the authors acknowledge were 'fixed using knowledge of where the zero-shot panel fails on the evaluation scenarios, not on a separate tuning set' (§9). The authors frame 0.990 as 'an optimistic upper bound,' which is the right characterization, but the abstract and conclusion still lead with 'parity without training' as the headline. The concern is bounded by the ablation (Table 6: few-shot alone reaches 0.983 on a harder seed set) and by cross-backend replication (Table 8: Gemini 0.715→0.995 with the same configuration), both of which suggest the result is not fragile to the exact lever combination. However, neither test addresses the core issue: the same tuned hyperparameters were applied to both backends and a

    Authors: We agree with the referee's core concern. The manuscript already acknowledges in §9 (Limitations) that the 0.990 figure is 'best read as a configuration tuned toward this eval, i.e., an optimistic upper bound,' and that 'a blind dev/eval split is the clean fix.' However, the abstract and conclusion do not adequately reflect this caveat — they lead with 'parity without training' as the headline, which overstates what the evidence supports. This is a fair criticism and we will revise accordingly. revision: partial

  2. Referee: (Continued from above — the referee's comment appears truncated, but the concern extends to: the same tuned hyperparameters were applied to both backends and all evaluation scenarios, so neither the cross-backend replication nor the ablation constitutes an independent test of the tuning procedure.

    Authors: The referee is correct that neither the cross-backend replication (Table 8) nor the ablation (Table 6) constitutes an independent test of the tuning procedure, since both reuse the same hyperparameters (k=6, N=3, τ=0.7, confusable-pair weighting). We acknowledge this explicitly in §9 but will strengthen the language in the abstract and conclusion to match. Specifically, we will make the following changes: (1) The abstract will replace 'parity without training' as the lead headline with a more measured statement: the enhanced panel reaches 0.990 classification accuracy — an optimistic upper bound under configuration tuned toward the evaluation set — compared to 0.985 for the supervised RF, with the gap closing on the same n=200 seeds. (2) The conclusion will foreground the zero-shot detection result (McNemar p=0.07, no significant difference) and the honest classification gap (0.685 vs. 0.985) as the primary findings, with the enhanced-panel result presented as a constructive but caveated secondary result. (3) We will add a sentence in both the abstract and conclusion noting that the hyperparameters were not selected on a held-out dev split and that the 0.990 figure should be read as an upper bound pending blind dev/eval validation. We will not remove the 0.990 result or the Table 5 comparison, as the ablation (few-shot alone reaching 0.983 on a harder, disjoint seed set) and the cross-backend replication (Gemini 0.715→0.995) do provide meaningful — if not fully independent — evidence that the effect is not fragile to the exact lever combination. The referee's framing of these as tests that 'suggest the result is not fragile' while not 'addressing the core issue' is precisely the characterization we will adopt in the revised text. We also note that the confusable-pair演示n revision: no

Circularity Check

0 steps flagged

No significant circularity; the few-shot demonstrations come from a disjoint seed range and the result is not forced by construction. The paper transparently acknowledges eval-set hyperparameter tuning, which is a generalization risk, not circularity.

full rationale

The paper's central claim is that curated contrastive few-shot demonstrations and self-consistency voting raise classification accuracy from 0.685 to 0.990, matching supervised ML. Walking the derivation chain: (1) The twin generates six fault classes, three engineered to be confusable on temperature (§3). (2) The zero-shot agent fails exactly on those confusable pairs (§6, Table 4). (3) Few-shot demonstrations over-weighting the confusable pairs are drawn from seeds 500–505, explicitly disjoint from both ML training seeds (0–299) and evaluation seeds (10000+), so there is no data leakage. (4) The enhanced agent classifies at 0.990 on the eval set. The paper does not claim the ICL mechanisms as novel ('We do not claim these ICL mechanisms as novel — contrastive few-shot and self-consistency are established techniques,' §1). No self-citation chain exists; all citations are to external work. The one methodological concern — that k=6, N=3, τ=0.7, and confusable-pair weighting were 'fixed using knowledge of where the zero-shot panel fails on the evaluation scenarios, not on a separate tuning set' (§9) — is a hyperparameter-overfitting risk that the paper itself flags as making 0.990 'an optimistic upper bound.' This is a generalization concern, not circularity: the demonstrations are not drawn from the eval set, the ablation on disjoint seeds (Table 6, n=60) shows few-shot alone reaching 0.983, and the cross-backend replication on Gemini (Table 8, 0.995) provides independent evidence the configuration is not eval-specific. The result is not equivalent to its inputs by construction. Score 2 reflects the acknowledged eval-tuning of hyperparameters, which is a minor methodological weakness but does not make the derivation circular.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 2 invented entities

The free parameters are mostly engineering constants tuned for telemetry realism and standard ICL hyperparameters. The main concern is that k=6, τ=0.7, and confusable-pair weighting are fixed using eval-set knowledge (acknowledged by authors). The axioms are domain assumptions about simulator fidelity, all explicitly flagged. The invented entities (twin, agent panel) are open-source and have some independent validation.

free parameters (10)
  • T² floor constants (ṅ₃=500 μmol/s, f=1.5, cold-plate Q₁₀₀=300 μW, still cooling 30 mW) = Tuned so unloaded MXC sits near 12 mK
    Engineering approximation to match BlueFors base state; not fitted to fault data but tuned for telemetry realism.
  • Quench transient parameters (τᵣ≈0.8 min, τd≈25 min, severity s) = τᵣ, τd hand-set; s varies per scenario
    Difference-of-exponentials pulse parameters chosen to produce realistic 4K flange excursion.
  • Few-shot demo count k=6 = 6
    Fixed using knowledge of zero-shot failures on eval set, not on a blind dev split (§9, Table 10).
  • Self-consistency sample count N=3 = 3
    Standard value from prior literature; not tuned on dev split.
  • Self-consistency diversity temperature τ=0.7 = 0.7
    Fixed without dev-split tuning (Table 10).
  • Realism stressor sev_scale=0.5 = 0.5
    Chosen to make ML model non-trivially perfect; justification for this specific value is not given.
  • Confusable-pair demonstration weighting = Over-weight helium_leak, blocked_impedance, wiring_heat_ingress
    Uses a priori physics knowledge of thermal degeneracy, but weighting is not formally specified or tuned on a dev split.
  • Noise covariance Σ (per-stage) = Estimated from BlueFors logs: MXC 0.74%, 50K 1.6%, flow 2.3%
    Learned from real data, not free in the traditional sense, but a fitted fingerprint.
  • Monitor poll cadence (30 min) = 30 min
    Fixed operational parameter; detection latency is cadence-bounded by this choice.
  • Monitor rolling window (4 h look-back) = 4 h
    Fixed operational parameter.
axioms (5)
  • domain assumption The T² dilution-cooling floor (Eq. 1) is a valid engineering approximation for stage temperatures in a dilution refrigerator.
    §3: 'the functional form is an engineering approximation sufficient for telemetry realism, not a first-principles derivation of the mixture thermodynamics.'
  • domain assumption Per-stage relative noise and cross-stage correlations estimated from one BlueFors log ('blizzard', 2021-10-08) generalize to other fridges.
    §3: noise fingerprint estimated from a single public log. The twin is distinguishable from real telemetry (AUC=1.00), so this is an acknowledged approximation.
  • domain assumption The six fault classes and their telemetry signatures (heat-load perturbations, flow/pressure signatures) faithfully represent real cryogenic faults.
    §3: faults are injected as heat-load perturbations. No real fault data validates the signatures; the sim-to-real check (§9) validates detection on real noise, not fault signatures.
  • domain assumption A compact numeric summary (per-channel start/end/%-change plus coarse trajectory) preserves sufficient information for diagnosis.
    §4: window presented as compact summary, not raw rows. Information loss is not quantified; the supervised ML path uses 120-d features from raw data.
  • domain assumption LLM model APIs (Claude Opus 4.8, Gemini 3.1 Pro) are stable enough for reproducible results.
    §10: model IDs are pinned, but proprietary APIs may change over time.
invented entities (2)
  • Onnes digital twin independent evidence
    purpose: Forward physics model with learned noise fingerprint that generates realistic dilution-fridge telemetry for benchmarking
    The twin is validated against real BlueFors logs for noise magnitudes and against two public cryostat corpora for base temperatures. A classifier two-sample test (AUC=1.00) honestly quantifies the fidelity gap. The code is open-source.
  • Five-role agent panel (Sentinel, Diagnostician, Operator, Guardian, Supervisor) independent evidence
    purpose: Multi-agent LLM pipeline for cryogenic fault diagnosis with separable safety veto and audit trail
    The ablation (Table 6) shows the panel matches a single-call agent on accuracy, and the Guardian veto rate (7.6-9.0%) is measured from logs. The panel's value is explicitly scoped to auditability, not accuracy.

pith-pipeline@v1.1.0-glm · 25750 in / 4204 out tokens · 449133 ms · 2026-07-08T23:30:22.143345+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 16 internal anchors

  1. [1]

    Many-Shot CoT-ICL: Making In-Context Learning Truly Learn

    T. T. Chung, L. Liu, M. Yu, and D.-Y . Yeung, “Many- Shot CoT-ICL: Making In-Context Learning Truly Learn,”arXiv:2605.13511, ICML 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Time-RA: Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback

    Y . Yang, Z. Liu, L. Song, K. Ying, Z. Wang, T. Bam- ford, S. Vyetrenko, J. Bian, and Q. Wen, “Time-RA: Towards Time Series Reasoning for Anomaly Diagno- sis with LLM Feedback,”arXiv:2507.15066, ACL 2026 Findings

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-Consistency Improves Chain of Thought Reasoning in Language Models,”arXiv:2203.11171, ICLR 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Self-Correction as Feedback Control: Error Dynamics, Stability Thresholds, and Prompt Interventions in LLMs

    A. Liu and J. Meng, “Self-Correction as Feedback Con- trol: Error Dynamics, Stability Thresholds, and Prompt Interventions in LLMs,”arXiv:2604.22273, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [5]

    Multi-Agent Design: Op- timizing Agents with Better Prompts and Topologies,

    H. Zhou, X. Wan, R. Sun, H. Palangi, S. Iqbal, I. Vuli´c, A. Korhonen, and S. Ö. Arık, “Multi-Agent Design: Op- timizing Agents with Better Prompts and Topologies,” arXiv:2502.02533, ICLR 2026

    work page arXiv 2026

  6. [6]

    The Illusion of Multi-Agent Advantage

    P. Jwalapuram, H. Lin, C. Li, F. Jiao, S. Wang, Y . Ming, Z. Ke, C. Qin, G. Carenini, and S. Joty, “The Illusion of Multi-Agent Advantage,”arXiv:2606.13003, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Delayed Verification Destabilizes Multi-Agent LLM Belief: Instability Thresholds and Optimal Corrector Placement

    I. Itkin, “Delayed Verification Destabilizes Multi-Agent LLM Belief: Instability Thresholds and Optimal Correc- tor Placement,”arXiv:2606.27409, 2026. 16

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

    L. Grinsztajn, K. Flöge, O. Key,et al., N. Hollmann, and F. Hutter (Prior Labs), “TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models,” arXiv:2511.08667, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Exploring Differences Between Tabular Enterprise Data and Public Benchmarks

    M. J. Kim, M. Schambach, F. Essenberger, A. Sres, and J. Höhne, “Exploring Differences Between Tabular Enter- prise Data and Public Benchmarks,”arXiv:2606.30452, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    Using LSTM Recurrent Neural Networks for Monitoring the LHC Su- perconducting Magnets,

    M. Wielgosz, A. Skocze´n, and M. Mertik, “Using LSTM Recurrent Neural Networks for Monitoring the LHC Su- perconducting Magnets,”Nucl. Instrum. Methods A, vol. 867, pp. 40–50, 2017

    work page 2017

  11. [11]

    Model Learning Algorithms for Anomaly Detection in CERN Control Systems,

    F. Tilaro, B. Bradu, M. Gonzalez-Berges,et al., “Model Learning Algorithms for Anomaly Detection in CERN Control Systems,” inProc. ICALEPCS, 2017

    work page 2017

  12. [12]

    Ma- chine Learning Framework for Anomaly Detection and Maintenance Optimization in Large-Scale Cryogenic Systems,

    P. Cacace, D. R. Santos, L. Giusti, F. Ferrand,et al., “Ma- chine Learning Framework for Anomaly Detection and Maintenance Optimization in Large-Scale Cryogenic Systems,”IOP Conf. Ser., 2025

    work page 2025

  13. [13]

    Super- conducting Radio-Frequency Cavity Fault Classification Using Machine Learning at Jefferson Laboratory,

    C. Tennant, A. Carpenter, T. Powers,et al., “Super- conducting Radio-Frequency Cavity Fault Classification Using Machine Learning at Jefferson Laboratory,”Phys. Rev. Accel. Beams, vol. 23, 114601, 2020

    work page 2020

  14. [14]

    cryometrics: BlueFors dilution- refrigerator log samples,

    L. Chen, “cryometrics: BlueFors dilution- refrigerator log samples,” GitHub repository, github.com/larchen/cryometrics. Fridge “blizzard,” logs dated 2021-10-08 (accessed 2026)

    work page 2021

  15. [15]

    Cryostat temperature logs: dataset for ‘Directed delivery of terahertz frequency radiation within a dry 3He dilution refrigerator’,

    University of Leeds Research Data Repository, “Cryostat temperature logs: dataset for ‘Directed delivery of terahertz frequency radiation within a dry 3He dilution refrigerator’,” CC BY 4.0, archive.researchdata.leeds.ac.uk/1034 (ac- cessed 2026). DR-200 dilution-fridge cool-down logs, Sep 2020–Mar 2022

    work page 2026

  16. [16]

    Process and control variables from the SNS’s cryogenic moderator system,

    B. Maldonado, D. Winder, P. Ramuhalli, and W. Blokland, “Process and control variables from the SNS’s cryogenic moderator system,” Oak Ridge National Laboratory [data set], 2024. doi:10.13139/OLCF/2441156

    work page doi:10.13139/olcf/2441156 2024

  17. [17]

    Enabling Technologies for Scalable Superconducting Quantum Computing,

    X. Croot, K. Nowrouzi, C. Spitzer, C. G. Almudever, A. Blais, M. Carroll,et al., “Enabling Technologies for Scalable Superconducting Quantum Computing,” arXiv:2512.15001, 2025

    work page arXiv 2025

  18. [18]

    Design and Operation of Wafer-Scale Packages Containing >500 Supercon- ducting Qubits,

    O. W. Kennedy, W. Ahmad, R. Armstrong, A. Awawdeh, A. Bose, K. G. Crawford,et al., “Design and Operation of Wafer-Scale Packages Containing >500 Supercon- ducting Qubits,”arXiv:2602.12773, 2026

    work page arXiv 2026

  19. [19]

    Integration and Resource Estimation of Cryoelectronics for Superconducting Fault-Tolerant Quantum Computers

    S. Kawabata, “Integration and Resource Estimation of Cryoelectronics for Superconducting Fault-Tolerant Quantum Computers,”arXiv:2601.03922, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  20. [20]

    A Cryogenic Hybrid Photonic/CMOS Controller Architecture for Scalable Superconducting Qubit Control

    B. Liu and Z. R. Huang, “A Cryogenic Hybrid Pho- tonic/CMOS Controller Architecture for Scalable Super- conducting Qubit Control,”arXiv:2606.10114, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  21. [21]

    A digitally controlled silicon quantum processing unit

    HRL Quantum Teamet al., “A digitally controlled sili- con quantum processing unit,”arXiv:2604.16216, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [22]

    Cryogenic Systems for Quantum Photonic Technologies: A Practical Review

    A. H. Rubin, V . A. Norman, and M. Radulaski, “Cryo- genic Systems for Quantum Photonic Technologies: A Practical Review,”arXiv:2605.12285, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    Vibrational sensing at mK temperatures in dry dilution refrigerators using commercial accelerometers for diverse fundamental physics applications

    N. Brace, A. D’Addabbo, S. D’Eramo, S. H. Fu, M. T. Hurst, T. O’Donnell,et al., “Vibrational sensing at mK temperatures in dry dilution refrigerators using com- mercial accelerometers for diverse fundamental physics applications,”arXiv:2601.08817, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  24. [24]

    Toward Full Autonomous Laboratory Instrumentation Control with Large Language Models

    Y . Xie, K. He, and A. Castellanos-Gomez, “Toward Full Autonomous Laboratory Instrumentation Control with Large Language Models,”arXiv:2604.03286, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Revisiting Classifier Two- Sample Tests,

    D. Lopez-Paz and M. Oquab, “Revisiting Classifier Two- Sample Tests,”International Conference on Learning Representations (ICLR), 2017

    work page 2017

  26. [26]

    Random Forests,

    L. Breiman, “Random Forests,”Machine Learning, vol. 45, no. 1, pp. 5–32, 2001

    work page 2001

  27. [27]

    LightGBM: A Highly Efficient Gradient Boosting Decision Tree,

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree,” inAdvances in Neu- ral Information Processing Systems (NeurIPS), vol. 30, 2017

    work page 2017

  28. [28]

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Temporal 2D-Variation Modeling for Gen- eral Time Series Analysis,”arXiv:2210.02186, ICLR 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy

    J. Xu, H. Wu, J. Wang, and M. Long, “Anomaly Trans- former: Time Series Anomaly Detection with Associa- tion Discrepancy,”arXiv:2110.02642, ICLR 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  30. [30]

    The Use of Confidence or Fiducial Limits Illustrated in the Case of the Bino- mial,

    C. J. Clopper and E. S. Pearson, “The Use of Confidence or Fiducial Limits Illustrated in the Case of the Bino- mial,”Biometrika, vol. 26, no. 4, pp. 404–413, 1934

    work page 1934

  31. [31]

    Note on the Sampling Error of the Dif- ference Between Correlated Proportions or Percentages,

    Q. McNemar, “Note on the Sampling Error of the Dif- ference Between Correlated Proportions or Percentages,” Psychometrika, vol. 12, no. 2, pp. 153–157, 1947. 17

    work page 1947