Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems

Advaith Nila Narayanan; Sriram Nagaraj

REVIEW 3 major objections 6 minor 12 references

Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Agentic AI can be governed by treating the inventory as a live graph in which a validation failure in one agent automatically blocks every agent that depends on it, directly or indirectly.

desk verdict A coherent and honest governance framework for agentic AI, but the matched-bootstrap calibration in §6.4 doesn't control the production false-positive rate as claimed. read the letter →

arxiv 2607.23916 v1 pith:776VR3MV submitted 2026-07-27 math.NA cs.NA

Sriram Nagaraj , Advaith Nila Narayanan This is my paper · ORCID

classification math.NAcs.NA

keywords modelriskmanagementagenticAIInventory-as-CodeDirectedAcyclicGraphpropagationchain-of-thoughtembeddingsdriftdetectionmatchedbootstrap

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that traditional model risk management, built around static registries and point-in-time validation, is structurally obsolete for agentic AI systems whose behavior is path-dependent and self-directed. It proposes replacing the ledger with an Inventory-as-Code feedback loop: the inventory is a directed acyclic graph of agents, a failure in one node propagates along dependency edges to block exactly its transitive descendants, and a materiality score is designed so tool breadth alone cannot inflate an agent's tier. For drift, it treats an agent's Chain-of-Thought as a monitored artifact and calibrates a cosine-drift threshold with a matched bootstrap that mirrors the production sample sizes, avoiding the false-positive inflation of naive half-split calibration. The contribution is a single, internally consistent governance loop with a containment guarantee that holds by construction rather than by calibration.

What carries the argument

The load-bearing objects are (i) the agent dependency DAG G=(V,E) with ancestor/descendant sets; (ii) the Composite Risk Score R_C_j and its blocking condition Eq. (8), which sets Status=Blocked iff any ancestor is Failed or Blocked, making blast radius equal the transitive descendant set by construction; (iii) the Degree of Autonomy materiality score R = α(I×A) + (1−α)T with T capped at T_max=20, so tool breadth alone can never lift an agent above Tier 3; and (iv) the Trajectory Drift Metric D_t = 1 − cos(g,p) comparing a certified Golden Path centroid against a rolling production centroid, with threshold δ calibrated by a matched bootstrap that draws N_val and N_prod samples per replicate.

What would settle it

On production or synthetic traces with temporal autocorrelation in in-distribution windows (e.g., slowly shifting market regimes), compute the empirical false-positive rate of the matched-bootstrap threshold at the stated α_fp; if it substantially exceeds α_fp, the exchangeability null is violated and the clean separation in §8.4 does not transfer. Also, an attempted adversarial input that preserves D_t < δ while altering reasoning semantics would falsify the monitor's claim to catch reasoning drift.

Watch

Extended reading notes

Core claim

The central claim is that agentic-system risk can be contained by three coupled mechanisms: a Degree of Autonomy materiality score with a capped tool-complexity term, a DAG-based Composite Risk Propagation rule whose blocking condition is pure reachability from any failed ancestor, and a Trajectory Monitoring protocol whose drift threshold is calibrated by a matched bootstrap that resamples a Golden Path and a production window at their actual sizes. The paper proves that the blast radius of a failure equals exactly the transitive descendant set of the failed agent, so quarantine is determined by graph contamination, not by an agent's inherent tier or by the penalty magnitudes. Simulation on

Load-bearing premise

The drift monitor's clean separation assumes production Chain-of-Thought embeddings are exchangeable with the certified validation corpus, so an i.i.d. matched bootstrap gives the true null distribution.

Editorial extensions

If this is right

If a root agent fails, every downstream consumer, no matter how many hops away, is automatically blocked, and no unrelated agent is affected; the quarantine decision is independent of penalty magnitudes.
Tool count alone cannot game an agent into a higher tier, closing a dominance loophole in additive risk scores.
The matched bootstrap restores false-positive control: in-distribution agents stay below threshold, while genuinely drifted agents are hard-flagged, under the exchangeability assumption.
Two-stage response separates a human-review state from irreversible quarantine, so single benign exceedances do not take agents offline.
LLM base-model version changes trigger automatic Golden Path invalidation and revalidation, preventing silent baseline decay.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The reachability-based blocking rule is not specific to agentic AI: any governed system whose components compose in a DAG (data pipelines, microservice dependencies, supply chains) inherits the same containment guarantee.
The matched-bootstrap calibration principle extends beyond cosine drift to any monitoring statistic whose null variance depends on two sample sizes; naive half-split schemes will systematically mis-set thresholds whenever N_prod differs from N_val.
The paper's own §9.1 concedes that production streams may violate exchangeability; a natural extension is a block-bootstrap or time-aware null, and a sensitivity analysis of α_fp under autocorrelated drift.
Adversarial trajectory manipulation—where an attacker keeps D_t below δ while shifting reasoning semantics—is flagged as an open problem; the framework as proposed detects statistical displacement of stated reasoning, not semantic equivalence.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Read this one if you work in model risk management for LLM systems. The paper proposes an Inventory-as-Code governance loop: a DoA materiality score with capped tool complexity, a DAG over agents with failure propagation, and a trajectory monitor that uses chain-of-thought embeddings to detect drift from a certified Golden Path. The framing is genuinely useful and the authors are unusually clear about what is and isn't established. The graph propagation result is correct but definitional—blocking is defined as reachability, so the blast radius theorem is true by construction. The DoA score is a sensible engineering answer to the tool-count dominance problem, and the two-stage soft/hard flag response is a practical improvement over a hard threshold. The paper also flags its own limitations, including autocorrelation in production streams and adversarial manipulation of CoT.

The soft spot is the calibration claim in §6.4. The paper says the matched bootstrap restores false-positive control by resampling both the Golden Path (N_val) and the production window (N_prod) from the validation corpus. But in production the Golden Path is fixed at registration; the only randomness in the drift metric under the null is the rolling window. By resampling g as well, the bootstrap null has variance proportional to 1/N_val + 1/N_prod instead of just 1/N_prod, so δ is inflated and the actual false-positive rate is below the advertised α_fp. The half-split criticism in the paper is valid—that null is indeed too tight—but it doesn't validate this two-sample resampling null. The simulation can't distinguish the two because a single seed with zero soft flags is consistent with both a correct 5% false-positive rate and a conservative one. This doesn't sink the framework, but the monitoring section needs a correct calibration: fix g and resample only p, or explicitly frame it as a two-sample test. The authors' own §9.1 acknowledges the i.i.d. bootstrap may be optimistic in real production, which is honest but doesn't cover this specific misspecification.

Citations are appropriate and the companion script makes the simulation reproducible. I'd send this to peer review—a competent referee will catch the calibration issue, and the revision should strengthen the paper. I'd cite the DAG propagation and DoA ideas, not the drift calibration.

Recommendation: engage with it for review, but expect the trajectory monitoring section to need rework.

Referee Report

3 major / 6 minor

Summary. The paper proposes an Inventory-as-Code governance loop for agentic AI systems, with four main contributions: a Degree of Autonomy (DoA) materiality score with a bounded tool-complexity term (Eq. 3), a DAG-based Composite Risk Propagation algorithm with a reachability-based blocking condition (Eq. 8), a trajectory drift monitor built on Golden Path centroids and a matched-bootstrap threshold (Eq. 11, §6.4), and practical protocols for LLM version changes, cycles, and prompt control. A synthetic 17-agent simulation (§8) is used to illustrate the framework: a single Tier 1 failure blocks exactly its transitive descendant set, and a matched-bootstrap calibration is claimed to cleanly separate in-distribution agents from drifted ones while the naive half-split null soft-flags roughly 90% of Tier 1 agents.

Significance. If the framework holds together, it provides a useful architectural template for extending model risk management to agentic systems: the graph-propagation logic is transparent, the DoA score separates tool breadth from impact/autonomy, and the trajectory-monitoring protocol tries to make drift thresholds a function of the production test statistic rather than an ad hoc cut. Strengths include explicit formulas, reproducible-simulation intent, and honest acknowledgement of limitations such as CoT unfaithfulness and non-i.i.d. production streams (§6.2, §9.1). However, the paper's central statistical claim about the matched bootstrap is not supported as written, the main 'theorem' is a restatement of the blocking definition, and the headline simulation is a single seed with no repeated-trial error analysis. The value is therefore more architectural than demonstrative at this stage.

major comments (3)

[§6.4, Def. 6.2 and Eq. (11)] The matched bootstrap resamples g* (size N_val) and p* (size N_prod) independently from the validation corpus. In production, however, D_t compares the registered, fixed Golden Path g_vi to a rolling production centroid p^(t). The correct null distribution is the distribution of D(g_vi, p*) conditional on the observed g_vi, whose variance scales as 1/N_prod. Resampling both centroids produces a null with variance scaling as 1/N_val + 1/N_prod. Since 1/N_val > 0, the threshold δ is inflated and the actual per-window false-positive rate is below α_fp. This contradicts the §6.4 claim that the matched bootstrap 'restores false-positive control at the intended α_fp'. The discrepancy is substantial in the paper's own parameter ranges (e.g., Tier 3 has N_val=50 and N_prod=100, doubling the variance scale). It also makes the Stage 2 threshold 1.5δ more conservative. The half-split comparison sho
[§8.4] The headline separation (10 Within Bounds, 0 Soft, 7 Hard) is one synthetic realization with a single seed and a single drift injection magnitude. The text itself notes that 'this seed realizes none' of the residual soft flags. With M=1000 bootstrap replicates but only one inventory configuration and one production realization, the simulation cannot provide evidence for 'false-positive control at the intended α_fp': a conservative calibration will also produce zero soft flags in one seed. Report repeated-seed distributions, or at least a Monte Carlo estimate of the achieved FPR under the null, and show the actual δ values under both the proposed two-sample bootstrap and a fixed-g conditional calibration.
[§5.3, Proposition 5.1] Proposition 5.1 is true but is a direct restatement of the blocking rule in Eq. (8): 'Blocked' is defined as having a contaminated ancestor, so the blocked set is exactly the union of transitive descendant sets by construction. The proof is a finite backward walk in a DAG. Since the paper itself says the guarantee 'holds by construction,' it should not be presented as a separate theorem or as an independent consequence of the composite risk formula. Reframe it as a design property of Eq. (8), and clarify that the β penalties carry no weight in the containment decision.

minor comments (6)

[§4.4, Remark 4.1] The claim that 'tool breadth alone can never lift an agent above Tier 3' is conditional on α=0.6 and T_max=20. The paper notes T_max is a design parameter, but a short sensitivity statement for other choices would prevent overgeneralization.
[§6.4] The term 'matched bootstrap' is misleading because g* and p* are drawn independently; consider 'size-matched two-sample bootstrap' or 'two-sample resampling null'.
[§8.4 / §8.1] The drift injection magnitude of 0.60 as an 'orthogonal component' is not formally defined in the setup. Specify how the synthetic CoT embeddings are transformed under drift (e.g., addition of an orthogonal vector followed by renormalization).
[§6.3] The certification criteria include a variance ceiling σ²_max and a 10% human-review requirement, but no procedure is given for setting σ²_max or auditing the review. Add operational guidance or cite a standard.
[General] The paper refers to a 'companion Python script' but provides no repository link, DOI, or checksum. Include an accessible reference for reproducibility.
[§6.4 vs §9.1] The main text states the matched bootstrap 'restores false-positive control at the intended α_fp' without qualification, while §9.1 concedes that the i.i.d. bootstrap null 'can be optimistic' under non-i.i.d. production. These statements should be reconciled; the main-text claim should carry the same caveat.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 'blast radius equals transitive descendant set' is a restatement of the blocking definition; the other framework components are independent, while the matched-bootstrap FPR claim is a correctness gap rather than a circularity.

self definitional [§5.3, Eq. (8), Proposition 5.1]
"Statusj =Blocked⇐ ⇒ ∃vi ∈ A(vj) with Status i ∈ {Failed,Blocked},(8) ... Proposition 5.1 (Blast radius equals reachability from a failure). Under Eq. (8), evaluated in any topological order, an agent vj that passes its own behavioral test is Blocked if and only if A(vj) contains a Failed agent. Equivalently, the set of Blocked agents is exactly the transitive descendant set D(vf ) of the failed agent(s) vf."

The proposition's conclusion is the definition in Eq. (8). 'Blocked' is defined as having a contaminated ancestor; in a DAG, the agents with a contaminated ancestor are, by definition, the transitive descendant set of the failed agent. The proof only unfolds this definition. The paper itself says the containment guarantee 'holds by construction,' confirming that the claimed result is a restatement of the chosen rule, not an independent derivation or prediction.

full rationale

Circularity is concentrated in Proposition 5.1. Eq. (8) defines Blocked as 'has a contaminated ancestor,' and Proposition 5.1 'proves' that the blocked set is exactly the transitive descendant set of a failure. In a DAG this is the same statement, so the headline propagation result adds no content beyond the rule selected. The paper is honest in calling this 'by construction,' but it is still a definitional equivalence presented as a formal contribution. The DoA 'tool breadth cannot lift tier' claim is likewise a consequence of the chosen constants alpha=0.6 and T_max=20, not an externally derived invariant; it is a calibrated design property rather than a circular prediction, and the paper notes T_max is a design parameter. No load-bearing self-citations appear; references are standard external works. The trajectory-monitoring section has a serious statistical concern: the deployed statistic uses a fixed Golden Path g_vi (Definition 6.1), while Section 6.4 calibrates delta by resampling both g* and p*, giving the calibration null extra variance and leaving the claimed alpha_fp control for a fixed baseline unestablished. This is a correctness/validation gap, not a circular Eq.-equals-Eq. reduction, and it is only partially acknowledged in Section 9.1's non-i.i.d. caveat. The simulation is self-contained but does not test the monitor against an external benchmark, so the clean separation in Section 8.4 is at best a sanity check. Overall, one central result reduces to its definition, giving partial circularity and a score of 6.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central framework depends on hand-selected calibration constants (α, T_max, βs, tier cutoffs, tool weights) and on strong domain assumptions about acyclicity, embedding stability, and exchangeability of production windows. The only introduced entities are procedural artifacts (Golden Path, Agency Risk, IaC loop), none with independent empirical evidence.

free parameters (6)

α (DoA blend weight) = 0.6
Chosen in §4.4 and used throughout; controls relative weight of impact-autonomy vs tool complexity and sets tier boundaries.
T_max (tool complexity cap) = 20
Design parameter in Eq. (1); the paper notes a real implementation would calibrate it against operational data. It makes 'tool breadth cannot lift tier' true for this choice.
β_F, β_B (propagation penalties) = 15.0, 5.0
Chosen in §5.2; affect the composite severity score but not the blocking decision because Eq. (8) is penalty-independent.
Tier boundary ratios (0.75 and 0.40 of R_max) = 0.75, 0.40
Arbitrary split points in Eq. (4) that determine Tier 1/2/3 assignment; no external calibration.
Tool severity weights w_s = 0.5, 1.0, 2.5, 5.0
Taxonomy in Table 1; chosen by the authors, not derived from data.
Simulation seed and drift injection magnitude = seed=42; 0.60 orthogonal
Single-seed synthetic setup in §8; the clean separation and 90% artifact depend on these choices.

assumptions (5)

domain assumption The agent dependency graph is a DAG; any cycles are handled by loop unrolling with a maximum iteration depth enforced by orchestration (§5.5).
If real agent graphs cannot be unrolled or have unbounded feedback, the topological propagation and blast-radius guarantee do not hold.
domain assumption CoT embeddings are computed by a stable, frozen embedding model ϕ and are L2-normalized (§6.2).
The Golden Path is defined in this embedding space; if ϕ changes, the baseline is invalid (the paper mandates suspension, §6.6).
domain assumption In-distribution production CoT windows are exchangeable with the validation corpus, making the i.i.d. matched bootstrap the correct null (§6.4).
The paper states this is needed for false-positive control and flags non-i.i.d. production as an open problem (§9.1).
domain assumption ValidationGuard behavioral tests are a trustworthy indicator of failure (§3.2, Algorithm 1).
All cascade behavior is driven by Failed labels from this gate; if tests miss failures, the propagation guarantees are irrelevant.
standard math Standard finiteness/acyclicity of reachability in finite DAGs (used in Prop 5.1's backward walk).
No issue beyond ordinary graph theory.

invented entities (3)

Golden Path (g_vi)
purpose: Reference centroid of validation CoT embeddings used as baseline for drift metric.
Constructed entirely from the validation corpus; no external handle for correctness beyond the certification criteria (which are themselves procedural).
Agency Risk
purpose: New risk category to cover non-deterministic reasoning, dynamic tool invocation, and inter-agent cascades.
A taxonomy label, not a measurable quantity with its own external operationalization; materiality is measured through the DoA score.
Inventory-as-Code control loop
purpose: Treats the inventory as active feedback system rather than a ledger.
Architectural concept; no direct falsifiable prediction, only the simulation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems." pith.science (2026). https://pith.science/paper/776VR3MV

@misc{pith2026260723916,
  author       = {Pith},
  title        = {Pith review of: Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/776VR3MV}},
  note         = {Machine review of arXiv:2607.23916}
}

read the original abstract

As financial institutions transition from traditional predictive models to autonomous agentic systems, the static model inventory requirements of traditional model risk management (MRM) face structural obsolescence. This paper proposes a dynamic Inventory-as-Code (IaC) governance loop that treats the model inventory as a living architectural component rather than a periodic documentation artifact. We make four principal contributions. First, we introduce a calibrated Degree of Autonomy (DoA) materiality score with an explicit, taxonomized tool-complexity weighting scheme that addresses the dominance problem of naive additive risk formulations. Second, we construct the agent inventory as a Directed Acyclic Graph (DAG) and define a formal Composite Risk Propagation algorithm under which upstream validation failures induce risk penalties on all reachable descendants. Third, we develop a Trajectory Monitoring protocol based on distributional cosine drift across ensembled Chain-of-Thought (CoT) embeddings, with an explicit procedure for constructing and certifying the Golden Path baseline, a matched-bootstrap calibration that we show is necessary to avoid a severe false-positive artifact in the naive alternative, and a two-stage response that separates legitimate reasoning variation from detrimental drift without demanding that a single exceedance event trigger an irreversible action. Fourth, we address practical complications largely absent from the prior literature: LLM base-model version changes, latent feedback loops in nominally acyclic agent graphs, and a two-pass execution structure that stages validation ahead of risk propagation.

Figures

Figures reproduced from arXiv: 2607.23916 by the authors.

**Figure 2.** Drift analysis simulation results: The figure shows per-agent trajectory [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 8 linked inside Pith

[1]

NIST AI 100-1, January 2023

National Institute of Standards and Technology.Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023. DOI: 10.6028/NIST.AI.100-1

doi:10.6028/nist.ai.100-1 2023
[2]

J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. arXiv:2201.11903

arXiv 2022
[3]

Turpin, J

M. Turpin, J. Michael, E. Perez, and S. R. Bowman. Language models don’t al- ways say what they think: Unfaithful explanations in chain-of-thought prompt- ing. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2305.04388

arXiv 2023
[4]

Lanham, A

T. Lanham, A. Chen, A. Radhakrishnan, et al. Measuring faithfulness in chain- of-thought reasoning. arXiv:2307.13702, 2023

arXiv 2023
[5]

Reimers and I

N. Reimers and I. Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProc. of EMNLP-IJCNLP, 2019. arXiv:1908.10084. 32

arXiv 2019
[6]

J. Gama, I. ˇZliobait˙ e, A. Bifet, M. Pechenizkiy, and A. Bouchachia. A survey on concept drift adaptation.ACM Computing Surveys, 46(4):Article 44, 2014

2014
[7]

Rabanser, S

S. Rabanser, S. G¨ unnemann, and Z. C. Lipton. Failing loudly: An empirical study of methods for detecting dataset shift. InAdvances in Neural Information Processing Systems (NeurIPS), 2019. arXiv:1810.11953

arXiv 2019
[8]

S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Confer- ence on Learning Representations (ICLR), 2023. arXiv:2210.03629

arXiv 2023

Show all 12 references

[9]

Schick, J

T. Schick, J. Dwivedi-Yu, R. Dess ` ı, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom. Toolformer: Language models can teach them- selves to use tools. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2302.04761

2023 arXiv
[10]

Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv:2308.08155, 2023

2023 arXiv
[11]

Efron and R

B. Efron and R. J. Tibshirani.An Introduction to the Bootstrap. Chapman & Hall/CRC, 1993

1993
[12]

M. T. Nygard.Release It!: Design and Deploy Production-Ready Software. Prag- matic Bookshelf, 2nd edition, 2018. 33

2018

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [1]

NIST AI 100-1, January 2023

National Institute of Standards and Technology.Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023. DOI: 10.6028/NIST.AI.100-1

doi:10.6028/nist.ai.100-1 2023

[2] [2]

J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. arXiv:2201.11903

arXiv 2022

[3] [3]

Turpin, J

M. Turpin, J. Michael, E. Perez, and S. R. Bowman. Language models don’t al- ways say what they think: Unfaithful explanations in chain-of-thought prompt- ing. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2305.04388

arXiv 2023

[4] [4]

Lanham, A

T. Lanham, A. Chen, A. Radhakrishnan, et al. Measuring faithfulness in chain- of-thought reasoning. arXiv:2307.13702, 2023

arXiv 2023

[5] [5]

Reimers and I

N. Reimers and I. Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProc. of EMNLP-IJCNLP, 2019. arXiv:1908.10084. 32

arXiv 2019

[6] [6]

J. Gama, I. ˇZliobait˙ e, A. Bifet, M. Pechenizkiy, and A. Bouchachia. A survey on concept drift adaptation.ACM Computing Surveys, 46(4):Article 44, 2014

2014

[7] [7]

Rabanser, S

S. Rabanser, S. G¨ unnemann, and Z. C. Lipton. Failing loudly: An empirical study of methods for detecting dataset shift. InAdvances in Neural Information Processing Systems (NeurIPS), 2019. arXiv:1810.11953

arXiv 2019

[8] [8]

S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Confer- ence on Learning Representations (ICLR), 2023. arXiv:2210.03629

arXiv 2023