acceptodds
Under review as a conference paper at ICLR 2027

Finding Icebergs in Language-Model Workflows: Diagnosing Latent Structural Fragility with Stochastic Semantic Evidence Graphs

Abstract

AI-workflow governance cannot be reduced to checking the final answer: an apparently safe answer may rest on a fragile evidence path that ordinary evaluation cannot see, localize or govern. We call this hidden fragility a "structural iceberg": hallucinations and unsupported claims may form its visible tip, while consequential weakness remains submerged. Stochastic semantic evidence graphs (SSEGs) expose these icebergs by preserving workflow channels, propagating local uncertainty and identifying the hidden paths on which an apparently safe output depends. ALCE and RAGTruth show that visible failures at the tip - unsupported citations and hallucinated spans - rest on distinct submerged weaknesses and therefore require different interventions. Across retrieval, tool-use and controlled stress tests, SSEG localizes those weaknesses, supports targeted repair, produces no false automatic passes in 35,000 known-truth cases and reduces ToolSandbox review by 28.8% across 96 executions from two agent models. The same structural view carries into end-to-end governance: in a separately sealed 1,200-case FinGovBench study, adding SSEG to GPT-OSS-20B reduces unsafe releases from 452/660 to 8/660 while releasing all 540 safe cases and correctly distinguishing 592/600 matched workflow pairs. An unchanged-gate transfer to Qwen3-8B releases all 540 safe cases and none of 660 unsafe cases, whereas flat-UQ releases 520 unsafe cases. SSEG therefore moves governance below surface-level output checking, turning hidden evidence dependencies into auditable, path-specific decisions about intervention, revalidation and release.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
AI governance, language models, agentic AI, semantic evidence graphs, uncertainty propagation, evidence attribution, retrieval-augmented generation, provenance, structural fragility, model evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_wqPjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper represents LLM workflows as stochastic semantic evidence graphs (SSEGs), recording dependencies between retrieval, generation, evaluation, and actions. A standard coupling argument propagates local discrepancies through edge influence bounds to bound terminal discrepancy. The proposed contribution connects this decomposition to review, revalidation, and release decisions. Validation includes 35,000 synthetic cases with known bounds; conflict-path localization on MAGIC; citation analysis and targeted review on ALCE, VitaminC, and ToolSandbox; and a sealed 1,200-case FinGovBench study with transfer to a second model. Reported results include improved localization, reduced review scope, and fewer unsafe releases, alongside null and negative results on additional tasks.

Critical strengths

The VitaminC and ToolSandbox results show that retaining dependencies can reduce the scope of revalidation or post-alarm review while preserving the reported decisions or alarm coverage. This is a useful practical benefit: source updates and detected failures can trigger targeted checks instead of requiring the entire workflow to be reviewed. The paper also reports null and negative results that help delimit these benefits.

Critical weaknesses
  1. The practical bounds are not established. Theorem 3.1 requires uniform local-discrepancy and edge-influence bounds. Although introduced alongside the definition of bvb_v, the inequality for cvpc_{vp} is an assumption, as the following sentence acknowledges. Algorithm 1 delegates construction of the bounds to methods named SimultaneousBound and InfluenceBound without specifying a procedure that establishes their validity. The paper does not specify the scoring and release rules sufficiently to reconstruct the real-task experiments: the FinGovBench release map, the flat-UQ rule, and the MAGIC score are not defined, and no values of bvb_v, cvpc_{vp}, or τ\tau are reported. Without a justified way to obtain these constants for real workflows, the theorem does not provide a usable certificate for the reported applications. Simulations where the bounds are known by construction cannot resolve this gap.

  2. The comparisons do not establish the added value of SSEG. FinGovBench includes paired cases in which the same local fault can affect the final action in one case but is blocked in the other. The correct decision differs between these cases, but the flat comparator does not receive the workflow connections needed to distinguish them. SSEG receives this information, so its advantage does not establish whether the proposed uncertainty propagation adds anything beyond access to those connections. A simple graph-aware baseline using the same local measurements, reachability, and hard controls seems to be missing. On MAGIC, SSEG uses both lexical scores and intervention measurements, while the comparator uses only lexical scores. The comparison therefore does not isolate the benefit of graph structure.

  3. The code does not reproduce the main decisions. The supplied repository's FinGovBench evaluation recomputes metrics from saved release decisions without regenerating those decisions from the model measurements. Appendix A and the README disclose this limitation, but it leaves the central step from measurements to decisions unverified. In addition, the available archive snapshot is dated after the ICLR deadline, and I cannot establish what code was available at that time. This limits both reconstruction of the reported experiment and verification of the submission's supporting implementation.

Questions for authors
  1. How are the bounds required by the theorem obtained and justified in the real-task experiments? Does any reported result use the resulting terminal bound?

  2. On FinGovBench, does SSEG outperform a simpler method given the same local measurements and workflow connections? On MAGIC, does the gain remain when the comparator also receives the intervention measurements?

  3. Can you provide code that regenerates the FinGovBench release decisions from the model measurements, and identify what code was available at the submission deadline?

Reviewer_d8y8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Stochastic Semantic Evidence Graphs (SSEG) to identify hidden structural weaknesses in LLM workflows. By modeling dependencies between evidence, intermediate components, and final decisions, SSEG propagates local uncertainty through a graph to estimate downstream discrepancies and identify critical paths for review. The authors evaluate the framework across synthetic stress tests, RAG benchmarks, tool-use workflows, and financial governance tasks, demonstrating improvements in risk detection, localization, and decision-making. Overall, this paper is well written.

Critical strengths

The paper addresses an important problem: reliable LLM evaluation should consider not only final outputs but also the underlying evidence dependencies. The proposed SSEG framework is interesting and provides a reasonable approach to modeling and propagating uncertainty across workflow components. I appreciate the attempt to connect theoretical uncertainty bounds with practical governance decisions. The experiments cover diverse settings and show encouraging results, particularly in identifying problematic evidence paths and reducing unsafe releases. Overall, the paper is well motivated and presents a coherent framework.

Critical weaknesses

While the SSEG framework is interesting and the graph-based representation is intuitive, the necessity of the specific DAG formulation is not sufficiently justified. I'm not sure if the DAG utilization can be used for some causality-like judgement; or to say, without DAG, will SSEG will be quite different. Recent studies also raised similar concerns. Although the paper compares against topology-free baselines, it remains unclear whether alternative dependency-aware representations or simpler propagation mechanisms could achieve similar results. Moreover, the relationship between recorded workflow dependencies and actual causal influence deserves further clarification.

Another concern is the practical validity of the uncertainty propagation mechanism. Although the proposed pathwise bound is mathematically motivated, its effectiveness depends on reliable estimates of local discrepancies and edge transmission coefficients. Obtaining valid bounds from finite observations in complex LLM workflows may be challenging. Furthermore, the resulting path contributions represent upper-bound contributions rather than actual causal or decision-risk effects. A comparison with data-driven influence estimation or alternative propagation methods would help justify this design.

Questions for authors

Why is the proposed DAG-based stochastic representation necessary compared with simpler dependency-aware approaches, particularly for iterative workflows with feedback loops?

How are the local discrepancy bounds and edge transmission coefficients reliably estimated in practice? What guarantees remain when these estimates are inaccurate?

How closely do the proposed pathwise contributions correspond to actual downstream decision impact? Could the authors compare them with intervention-based or learned influence estimates?

Reviewer_cUTtOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces Stochastic Semantic Evidence Graphs (SSEGs) to detect hidden weaknesses in LLM workflows by tracking evidence dependencies and uncertainty. It also presents the ELRG framework and evaluates the approach on several benchmarks.

Critical strengths
  1. The paper addresses an interesting and practically relevant problem: identifying hidden risks in LLM workflows beyond evaluating final outputs.
  2. The proposed framework provides a structured approach to uncertainty propagation and evidence-path localization.
  3. The paper offers a theoretical formulation to support the proposed approach.
Critical weaknesses
  1. Limited experiments or empirical evidence. The main paper provides limited results, with key evidence deferred to the appendix. Table 2 evaluates only two LLMs. More experiments with SOTA baselines and ablations are needed to support the claims.
  2. Poor presentation. The paper lacks clear organization and ideas are not well-delivered with logic or coherency. For example, the Introduction is overly long and incoherent; also Figure 1's caption is too long to understand
Questions for authors
  1. Ablation studies: Can the authors test the effect of removing graph topology, path influence coefficients, and intervention information one at a time while keeping local measurements fixed?
  2. Stronger baselines: Can SSEG be compared with methods that use workflow traces, dependency structures, or graph-based risk localization?
  3. End-to-end evaluation: Can the authors test SSEG on realistic RAG and tool-use workflows, measuring error detection, localization accuracy, review cost, and final decision quality?
  4. More Tested LLMs: Can the authors include more models from different families and sizes in Table 2 to show SSEG's generalizability?
  5. Model qualification failures: Why did Mistral-7B and OLMo-2-7B fail the initial qualification? What does this mean for SSEG's applicability? Can the authors explain the failure and also test more models that meet the requirements?
Reviewer_wfQSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how a change in a source, tool or model output can affect a later decision in an AI workflow. Its evidence graph records which components depend on each other. Given bounds on local changes and how far each change can travel, Theorem 3.1 bounds the change in the final decision distribution relative to a reference workflow. The graph also identifies which sources to inspect after an alarm. The authors test propagation on simulated graphs, localization on labeled retrieval tasks, and release decisions on 600 matched FinGovBench pairs. Adding the graph-based rule to GPT-OSS reduced unsafe releases from 452 to 8 out of 660 and released all 540 eligible cases. Applied to Qwen on the same cases, the frozen rule released all 540 eligible cases and no unsafe ones.

Critical strengths
  1. The graph makes a warning actionable: it identifies which claims depend on a changed source and need another check. In the VitaminC replay, the method found the affected paths for all 688 revisions. It sent 688 claims through review, compared with 2,528 under exhaustive review, with the same release decisions from the entailment checker. This gives a concrete use for retaining dependencies even when an existing checker makes the final decision.
  2. The localization study tests the ranking against independently labeled conflicts. On 960 held-out MAGIC cases, the lexical baseline ranked a conflicting path first in 75.8% of cases; the proposed score reached 80.3 to 82.3% across five models. On HotpotQA, the method improved the shortlist while reducing the chance that its very first choice was correct. That distinction helps a practitioner judge whether it fits a review queue or a single automatic choice.
Critical weaknesses

The main release test leaves the value of the probability scores unclear. In each of the 600 pairs, the same event affects the same step. What changes is whether that step feeds the final decision. For example, if a source is out of date, it matters whether the decision uses that source. The full method correctly handled both cases in 592 pairs with GPT-OSS (Table 2). The comparison rule receives the same measurements about each step, but ignores the connections between steps. That is a useful test of whether the connections matter. The table leaves out a simpler comparison: follow the connections, hold a decision that depends on the affected step, and honor any mandatory stop. Without that comparison, it is hard to tell whether a team needs the probability scores to make these release decisions.

Questions for authors

na

Reviewer_59PBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models a language model workflow as a DAG. Every node gets a local discrepancy b_v and every edge a transmission coefficient c_vp, and Theorem 3.1 bounds the discrepancy at the terminal action by the sum of local discrepancies times path influence. On top of this they run a four stage procedure, Expose, Localize, Route, Govern, which either releases the action or sends a named path for review. The evidence is a ladder: 35,000 synthetic cases for bound coverage, MAGIC for ranking the conflicting path, ALCE and VitaminC for routing review to a claim source edge, and a sealed FinGovBench study with GPT-OSS-20B plus an unchanged gate transfer to Qwen3-8B for the release decision. The appendices add Self-RAG, ToolSandbox, RAGTruth and a financial 2 by 2 study, several of which are reported as tied or failed.

Critical strengths

The problem is real. In a regulated workflow the question is not whether the final answer looks fine but which retrieval or tool step it depends on and what has to be rechecked when that step changes, and I have not seen this written as a path bound with a review action attached to it. Also the authors are careful in a way most papers in this space are not. Splits are held out before labels are opened, FinGovBench predictions are hashed before labels are evaluated (Appendix E), and the negatives stay in the paper: Mistral and OLMo failing the interface check (Section 4), the Self-RAG exact hop result being null at 38.6 versus 38.4 (Appendix G), the CONFLICTS pilot being pulled for a scoring bug. So this is good.

Critical weaknesses

The FinGovBench numbers in the abstract are mostly reading a label which is defined by the topology. In 540 of the 600 pairs the event and the local node descriptions are held fixed and the only thing which changes is whether the affected node is connected to the governed action or blocked from it (Appendix E). So the 540 safe cases are basically the blocked arms. SSEG releases all of them because the path influence of the defect is zero, and flat-UQ is defined to throw the topology away (Section 4), which is why exact pairs sits between 10% and 20.5% for every model only and flat-UQ row in Table 2. The Qwen transfer makes it clearest. Qwen alone releases 0 of 540 safe cases, so it is reporting a defect nearly everywhere, and the same frozen gate then gets 540/540 and 0/660, and only the topology can produce that. The paper itself makes this argument against the terminal drift comparator in Appendix G, Finding 3, that a score which reads the thing which defines the label is not an independent comparator. The same thing applies here. On top of this the benchmark is a 1.0.0-rc4 release candidate from the same author as [46], and Appendix E says a second independently sealed population is future work. I would like to see a reachability only rule, release when no defect node has a path to the action and withhold otherwise, with no uncertainty propagation at all. If it matches Table 2, the abstract should say so.

Questions for authors

Please run the reachability only gate from Weakness 1 on both models and put safe recall, unsafe release and exact pairs next to Table 2. If propagation adds something over reachability, show where

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.