acceptodds
Under review as a conference paper at ICLR 2027

From Clues to Causes: Teacher-Guided Reasoning and Analysis through Cascaded Evidence for Wafer Defect Map Diagnosis

Abstract

A complete wafer defect diagnosis should answer five questions in sequence: where the defect is, what it looks like, what type it is, why it occurred, and how it can be prevented. Existing wafer defect analysis methods mainly focus on defect localization, pattern recognition, and type classification, while root-cause diagnosis and actionable prevention remain largely unexplored. This leaves a substantial gap between visual defect recognition and process-level decision support. We propose TRACE, Teacher-guided Reasoning and Analysis through Cascaded Evidence, a reinforcement-learning framework that formulates wafer defect diagnosis as a five-stage reasoning process spanning defect position, appearance, type, root cause, and prevention. TRACE verifies the five diagnostic fields sequentially and halts at the first failed stage, preventing downstream rewards from being assigned on top of an already incorrect premise. A frozen snapshot of the policy holding the oracle then provides two forms of stage-specific guidance, neither of which directly reveals the answer. The first is a natural-language clue that constrains the reasoning space, while the second is a dense guidance signal derived from the centered difference between clue-conditioned and unconditioned logits of the same model. We show that this difference corresponds to a per-token pointwise mutual-information signal with the withheld clue information, and use it only to rescale the projection of the model’s own verifier-driven policy gradient. Because this rescaling acts as a positive-definite preconditioner rather than an auxiliary loss, it leaves the token-level stationary points of the original reinforcement-learning objective unchanged. No teacher, oracle, or clue is required to produce a diagnosis at inference time. We evaluate TRACE on a real-world industry wafer defect dataset using complete diagnostic-chain accuracy and counterfactual root-cause reasoning, where visually similar defect patterns arise from different manufacturing causes. The results demonstrate that TRACE improves not only defect recognition but also end-to-end diagnostic reasoning from visual evidence to root cause and prevention.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Wafer Defect Diagnosis, Vision-Language Models, Reinforcement Learning, Teacher-Guided Reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zC19OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses complete wafer defect diagnosis using per-operation adder maps (inspections before and after a single manufacturing step) alongside process context, moving beyond standard end-of-line pattern classification to infer five sequential diagnostic stages: key-position (S0S0), appearance (S1S1), defect type (S2S2), root cause (S3S3), and prevention (S4S4).

Critical strengths

A central challenge in RL for reasoning with privileged hints/teachers is that matching a privileged teacher via KL divergence or SFT (e.g., on-policy self-distillation) shifts the optimization target toward an oracle-conditioned distribution, suppressing exploration and failing to transfer cleanly to unhinted test-time inference.

The paper pinpoints a fundamental limitation in existing industrial anomaly/wafer benchmarks: end-of-line bin maps accumulate defects across all steps, making root-cause inference ill-posed, whereas naive outcome-level RL rewards models for guessing high-prior root causes even when visual localization fails. By pairing per-operation adder maps with cascaded verification (halting both reward accumulation and directional gradient rescaling for stages s>Fs > F) and evaluating on CF-RC (counterfactual pairs sharing identical process contexts and defect classes but differing in physical morphology and mechanism), the authors convincingly demonstrate that the model learns visual-to-causal grounding rather than process-table lookup (69.3% vs. 47.8% for an oracle-type rule lookup system).

Critical weaknesses

The core experimental data comes from internal production records of a specific semiconductor fab. Complete adder maps, process context and expert annotations cannot be fully disclosed. This prevents external researchers from reproducing the core experiments and makes it difficult to verify the method’s generalizability across other process nodes, fabs, or industrial scenarios. For academic research, the inaccessibility of the dataset significantly reduces the work’s replicability and extendability.

TRACE integrates multiple innovative components simultaneously: cascaded reward truncation, constrained clue retry, premise-propagation discount, and logit-shift gradient amplification. Although the paper performs partial ablations, interaction effects between components are not thoroughly discussed. For example, it remains unclear whether the gains from the clue channel and direction channel are additive or offsetting, and whether the cascade mechanism is redundant with GRPO’s inherent advantages. Furthermore, the division and ordering of the five stages are largely based on domain experience, and their necessity and optimality lack quantitative support.

Questions for authors

Could the authors provide the core specifications deferred to Appendices B, F, H, and I?

In Proposition 1, zθ(st∣h)z_\theta(s_t\vert{}h) conditions on clue hh in the prompt prefix, whereas log⁡pθ(h∣st,yt=v)\log p_\theta(h \vert{} s_t, y_t = v) treats hh as conditioned on the next token vv. Under an autoregressive parameterization where hh precedes sts_t, how is this exact equality derived in Appendix A without remainder terms from the softmax partition functions Z(st∣h)Z(s_t\vert{}h) and Z(st)Z(s_t)? Does mean-centering C(⋅)\mathcal{C}(\cdot) over the full vocabulary V\mathcal{V} before applying the top-10 projection Π10\Pi_{10} leave a non-zero mean on the retained 10 coordinates, and how sensitive is TRACE to the choice of k=10k=10 in Πk\Pi_k?

Stages S1,S3,S1, S3, and S4S4 are graded by a frozen snapshot of the student policy πˉθ\bar{\pi}_\theta refreshed every NN steps via P(Yes)/(P(Yes)+P(No))P(\text{Yes})/(P(\text{Yes})+P(\text{No})). How do you prevent self-reinforcing confirmation bias or stylistic reward hacking as πˉθ\bar{\pi}_\theta co-evolves with πθ\pi_\theta, especially in S4S4 where aconsa_{\text{cons}} is evaluated without oracle access?

Reviewer_1yCJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TRACE to address the gap of root cause and prevention in the wafer defect diagnosis process. The motivation is clear and solid. Building on model distillation, the proposed approach combines cascaded verification with gradient preconditioning on the clues. The experiment results show substantial improvements over SFT, GRPO, and distillation alternatives. However, there are some major concerns on the evaluation of the approaches and minor concerns on the reproduction of the approaches.

Critical strengths
  1. Clear motivation concerning the root cause linking visual evidence to the dianogisis process.
  2. Extensive comparisons among related baselines.
  3. Deployment doesn't require teacher model inference.
Critical weaknesses
  1. The rationale behind checking appearance before type indicating visual evidence usage is unclear.
  2. Evaluation grader seems to be model-based.
  3. Lack of reproducibility details.
Questions for authors
  1. Is the test-time grader identical across all methods, and is it independent of TRACE’s training trajectory?
  2. Why does checking appearance before type establish that the model uses visual evidence internally? Could a model recover the likely cause from process context while separately generating an acceptable appearance description?
  3. How do we verify the requirement that a clue remain informative but non-identifying?
  4. More detailed info about the benchmark and implementation could be reported for reproducibility.
Reviewer_mH6NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper address interactive wafer defect diagnosis at a process step, given process step attributes, process history, and defects attributed to that particular step. The framework appears to do a chain of inference to identify how the defect may be caused in that step and potential actions to fix the problem.

The authors map this task to an RL problem. The proposed framework is designed to provide five diagnostic answers. They are evaluated in a sequential manner. While the system appears to have many interesting elements in the specific semiconductor diagnosis setting, the paper lacks a coherent problem description as RL and it is difficult to assess its technical value.

Critical strengths
  • The task addressed appears to be more general than tradition wafer map defect classification and can be viewed as a first-of-a-kind application of the RL approach.
  • Designs a specific reward function that reflect specific requirements in semiconductor manufacturing.
  • RL helps output more relevant answer by adjusting internal workings of the policy function.
Critical weaknesses
  • The paper lacks a coherent formulation of the problem within the standard RL framework. The mapping between the semiconductor diagnosis task and the key RL components (state, action, transition, and reward) is not clearly described, making it difficult to understand how the proposed approach fundamentally differs from conventional reasoning or retrieval-based systems.
  • In addition, the paper provides very limited information regarding data availability, observability, and training data coverage. As a result, it is difficult to assess the feasibility of the proposed approach. In semiconductor manufacturing, historical diagnostic data are often sparse and incomplete, and it is unclear whether sufficient interaction data exist to effectively train an RL agent.
  • The proposed diagnostic chain appears to focus on a process step that has already been identified as responsible for the defect. This assumption may not reflect many real-world diagnosis scenarios, where the root-cause process step itself is unknown and must first be identified. If the responsible process step is given a priori, a simpler and more lightweight LLM-based retrieval or reasoning system may be sufficient, reducing the motivation for adopting an RL framework.
Questions for authors

Please comment on the points raised in the weakness section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.