acceptodds
Under review as a conference paper at ICLR 2027

Circuits of LLM Classifiers

Abstract

Language models make it easy to build classifiers by simply prompting a model with the sample to classify. However, their predictions are neither deterministic nor uniformly reliable across samples. How can we combine predictions from different models to improve reliability? We view LLM classifiers as *noisy measurement instruments* of an underlying target label. Our goal is to maximize average classification accuracy at a given average query cost. To this end, we introduce *circuits* that organize LLM classifiers into an adaptive sequence of queries and decisions. We then characterize the optimal circuit, obtain the fundamental limits of achievable accuracy, and provide practical insights on realizing the optimal circuit. Across safety, healthcare, political science and preference benchmarks, our method improves the cost–accuracy frontier over routing and cascading baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Classification, Sequential decision making, LLM circuits

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TrbrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LMCircuit treates LLM classifiers as imperfect tool for predicting labels and tries to combine their answers intelligently. It uses dynamic programming to decide which models to query and when to balance accuracy against cost. The authors provide theoretical results on how accuracy improves with additional information and show that LMCircuit outperforms several existing routing, cascading and voting methods across four benchmarks while remaining fast to train.

Critical strengths
  1. LMCircuit combines model selection, query ordering and stopping into one pricipled framework.
  2. It trains quickly and consistenly outperforms ten baselines across four diverse tasks.
  3. Paper explains why the method works, showing that is prioritizes models by information gained per unit cost and spends more queries when predictions are uncertain.
  4. The pool construction study offers practical guidance for selecting useful models.
Critical weaknesses
  1. Scaling-law analysis assumes instruments perform better than random guessing on every prompt but a substantial fraction of prompts in some benchmarks violate this condition. The theory’s applicability to these cases remains unclear.
  2. LMCircuit includes a supervised text-based classifier that can predict labels without asking any LLM but the paper does not report its standalone performance or compare against strong supervised stacking baselines under matched conditions.
Questions for authors
  1. Scaling-law analysis assumes that every instrument performs at least as well as random guessing on every prompt. How often is this assumption violated in the benchmarks and how robust are the theoretical predictions when such violations occur?
  2. How much of LMCircuit's improvement comes from its zero-cost classifier versus adaptive querying? Compare its standalone performance with the a supervised stacker under matched conditions to help isolate the contribution of adaptive querying.
3 comments

Thank you for your thoughtful and detailed review! We are glad that you appreciated LMCircuit's principled framework, its consistent improvements across various baselines, and the theoretical insights explaining its performance. Following your suggestions, we added the new results to the paper in Section C.4.4 of the appendix. We welcome more questions or suggestions.


Q1: Scaling-law analysis assumes that every instrument performs at least as well as random guessing on every prompt. How often is this assumption violated in the benchmarks and how robust are the theoretical predictions when such violations occur?

For the theoretical analysis, the assumption ai(z)≥1/2a_i(z) \geq 1/2 is without any loss of generality since an instrument with accuracy ai(z)<1/2a_i(z) < 1/2 provides the same information, denoted by Ii(z)I_i(z) in Line 412, as one with accuracy 1−ai(z)1-a_i(z) after reversing its answers. Thus, the results also apply to instruments that perform worse than random guessing, provided their accuracies are known. The purpose of these results is to characterize the fundamental limits of the accuracy–cost tradeoff under known instrument accuracies, establishing what is achievable with optimal use of the available instruments. We added footnote 5 addressing this point in Line 431.


Q2: How much of LMCircuit's improvement comes from its zero-cost classifier versus adaptive querying?

The standalone accuracy of LMCircuit's zero-cost classifier corresponds to the leftmost point of each frontier in Fig. 4: 79.85% on BeaverTails and 76.53% on PolitiCAUSE. To isolate the benefit of adaptive querying, we compare how much LMCircuit must spend to achieve the same target accuracy against two non-adaptive baselines.

  1. Supervised stacker. We train a text-only logistic regression on the prompt embeddings, followed by a separate logistic regression for each fixed subset of classifiers. The second model combines the text-only prediction with the selected classifiers' answers.

  2. Non-adaptive circuit. We use LMCircuit's own learned belief model to combine the answers of each fixed subset of classifiers. This baseline shares the same text-only prediction, belief model, and calibration as LMCircuit. The only difference is that it queries a fixed subset, whereas LMCircuit chooses which classifiers to query and when to stop based on the prompt and previously observed answers.

For each method, we select the lowest-cost operating point that achieves a given target accuracy on the validation set, considering all 2∣H∣2^{|\mathcal H|} fixed subsets for the baselines. We then report the resulting test costs and LMCircuit's percentage savings over each baseline, given by 100(1−CLM/Cbaseline)%100(1-C_{\mathrm{LM}}/C_{\mathrm{baseline}})\% (higher is better). Positive values indicate cost reductions (e.g., 50% means LMCircuit uses half the baseline's cost), while negative values indicate higher costs.

BeaverTails

Target accuracy (validation)LMCircuit costStacker costSavings vs. stacker [95% CI]Non-adaptive costSavings vs. non-adaptive [95% CI]
79%0.0060.02171% [64, 78]0.03583% [80, 86]
80%0.0210.04149% [43, 55]0.06367% [60, 71]
81%0.0490.06120% [3, 38]0.13163% [49, 72]
82%0.1050.18443% [25, 60]0.22152% [37, 66]

PolitiCAUSE

Target accuracy (validation)LMCircuit costStacker costSavings vs. stacker [95% CI]Non-adaptive costSavings vs. non-adaptive [95% CI]
78%0.0070.003−136% [−193, −63]0.03780% [75, 86]
79%0.0240.04649% [32, 65]0.05960% [47, 73]
80%0.0790.15549% [36, 62]0.19359% [49, 68]
81%0.4351.39169% [61, 76]0.77144% [26, 57]

Across both datasets, LMCircuit achieves substantial cost savings at most target accuracies. On BeaverTails, it reduces query costs by 20–71% relative to supervised stacking and 52–83% relative to the non-adaptive circuit. On PolitiCAUSE, the corresponding savings are 49–69% and 44–60% for targets of 79–81%. At the lowest PolitiCAUSE target, the stacker is cheaper. The cost savings therefore demonstrate the benefit of adaptively selecting classifiers and deciding when to stop. We added these results to Section C.4.4 (Line 1538) in the appendix.

Thank you for additional experiments. The ablations address my concerns about adaptive querying and the zero-cost prior. I would appreciate (a) similar results for LMArena and HealthBench, (b) test accuracy at each operating point and (c) details on CI computation.

Regarding Q1, I agree with the flipping argument when accuracies are known. My remaining concern is robustness to misestimated per-prompt accuracies, particularly on systematically incorrect prompts. A brief mention in the limitations would suffice.

Thank you for the follow-up! We provide the additional experiments and statistical details below and will include them in Section C.4.4.


Q1. Yes, these theoretical results assume known instrument accuracies and do not address errors in estimating prompt-specific reliability. We will explicitly acknowledge this limitation in the paper.


Q2.

(a–b) Additional comparisons and test accuracy. The tables below report 95% confidence intervals for both test accuracy and relative cost savings.

BeaverTails

Target (val)LMCircuit costLMCircuit test accStacker costStacker test accSavings vs. stackerNon-adaptive costNon-adaptive test accSavings vs. non-adaptive
79%0.00680.70% [79.27, 82.08]0.02180.66% [79.28, 81.95]71% [64, 78]0.03580.72% [79.34, 82.07]83% [80, 86]
80%0.02181.39% [79.97, 82.77]0.04181.02% [79.64, 82.31]49% [43, 55]0.06381.33% [79.96, 82.72]67% [60, 71]
81%0.04982.05% [80.73, 83.42]0.06181.39% [79.95, 82.75]20% [3, 38]0.13181.94% [80.64, 83.28]63% [49, 72]
82%0.10582.73% [81.43, 84.01]0.18482.00% [80.66, 83.30]43% [25, 60]0.22182.57% [81.28, 83.86]52% [37, 66]

PolitiCAUSE

Target (val)LMCircuit costLMCircuit test accStacker costStacker test accSavings vs. stackerNon-adaptive costNon-adaptive test accSavings vs. non-adaptive
78%0.00777.60% [76.13, 79.09]0.00376.48% [74.90, 78.02]−136% [−193, −63]0.03777.71% [76.35, 79.15]80% [75, 86]
79%0.02478.09% [76.59, 79.60]0.04677.66% [76.27, 79.13]49% [32, 65]0.05978.42% [76.97, 79.90]60% [47, 73]
80%0.07978.66% [77.19, 80.10]0.15578.53% [77.11, 80.02]49% [36, 62]0.19378.98% [77.60, 80.44]59% [49, 68]
81%0.43580.11% [78.71, 81.48]1.39179.89% [78.42, 81.36]69% [61, 76]0.77180.16% [78.75, 81.51]44% [26, 57]

LMArena

Target (val)LMCircuit costLMCircuit test accStacker costStacker test accSavings vs. stackerNon-adaptive costNon-adaptive test accSavings vs. non-adaptive
65%0.00464.46% [63.46, 65.53]0.07164.78% [63.87, 65.71]95% [90, 99]0.03364.73% [63.75, 65.78]89% [86, 94]
66%0.01364.83% [63.83, 65.92]0.13765.71% [64.81, 66.63]90% [86, 93]0.09565.52% [64.61, 66.52]86% [81, 90]
67%0.07066.15% [65.19, 67.16]0.28266.24% [65.28, 67.18]75% [71, 79]0.15766.32% [65.35, 67.32]56% [49, 62]
67.5%0.13066.69% [65.74, 67.67]0.72966.70% [65.71, 67.67]82% [79, 86]0.24566.66% [65.71, 67.66]47% [33, 58]

HealthBench

Target (val)LMCircuit costLMCircuit test accStacker costStacker test accSavings vs. stackerNon-adaptive costNon-adaptive test accSavings vs. non-adaptive
79.5%0.00379.93% [78.81, 80.96]0.12479.86% [78.83, 80.88]98% [97, 99]0.04579.88% [78.78, 80.88]94% [93, 96]
80%0.00880.26% [79.16, 81.26]0.32480.12% [79.08, 81.10]97% [97, 98]0.10180.10% [79.03, 81.09]92% [90, 93]

For completeness, we also report the highest test-accuracy achievable by each method, with their cost.

Highest-accuracy operating points

DatasetLMCircuit test acc / costStacker test acc / costNon-adaptive test acc / cost
BeaverTails83.46% / 0.7783.19% / 2.5683.61% / 2.37
PolitiCAUSE80.44% / 1.3779.89% / 2.4480.50% / 2.19
LMArena67.10% / 0.2966.76% / 1.6967.00% / 0.81
HealthBench81.05% / 0.6780.82% / 1.4980.88% / 2.33

Discussion. LMCircuit achieves lower query costs than the non-adaptive circuit at every reported target. Since both methods use the same learned belief model, these comparisons directly measure the benefit of adaptive classifier selection and stopping. The savings over supervised stacking are also substantial, except at the lowest PolitiCAUSE target, where the stacker is cheaper but has lower test accuracy.

(c) Confidence intervals. We compute 95% confidence intervals using 2,000 bootstrap resamples of the test set. For each resample, we compute test accuracy and relative cost savings, defined as 1−CˉLMCircuit/Cˉbaseline1-\bar C_{\mathrm{LMCircuit}}/\bar C_{\mathrm{baseline}}, and report the mean cost saving. The reported 95% intervals are the 2.5th and 97.5th percentiles of these estimates. For each prompt, we average cost and correctness over eight independent draws of classifier responses. We use the same resampled test prompts for LMCircuit and each baseline. We also retrain the belief model used by LMCircuit and the non-adaptive circuit five times, and report the average performance.

Reviewer_XmA2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces LMCircuit, a decision-theoretic framework for adaptively combining LLM classifiers. It formulates classifier selection and stopping as a dynamic program, derives theoretical cost–accuracy limits and scaling laws, and learns the required distributions using prompt embeddings and classifier observations. Experiments on four benchmarks demonstrate improvements over routing, cascading, and voting baselines.

Critical strengths
  1. Principled formulation: The dynamic programming framework unifies routing, cascading, and aggregation while providing a theoretically motivated adaptive querying strategy.

  2. Strong empirical validation: Experiments across four diverse tasks demonstrate consistent improvements in the cost–accuracy tradeoff against numerous baselines.

Critical weaknesses
  1. Distributional consistency: The label posterior and measurement distributions are estimated using separate heads without enforcing probabilistic consistency. This could undermine the optimality of the learned dynamic program under model misspecification.

  2. Cost accounting and scalability: Embedding and decision-making overheads are excluded from query costs, potentially overstating low-budget advantages. The exponential complexity also limits scalability beyond small classifier pools.

Questions for authors
  1. How sensitive is the learned policy to inconsistencies between the estimated posterior and measurement distributions?

  2. How does the cost–accuracy frontier change when embedding and inference overheads are included?

1 comment

Thank you for your thoughtful feedback! We are glad that you appreciated the paper's "principled formulation" and "strong empirical validation." Below, we address your questions about distributional consistency and computational overheads, and we added these new results to the paper (Appendix Section C.4). We welcome any further questions or feedback.


Q1: How sensitive is the learned policy to inconsistencies between the estimated posterior and measurement distributions?

As you correctly point out, the label posterior and measurement probabilities come from a single shared model but are trained through separate losses (Eqs. 47-48), without explicitly enforcing that they correspond to a consistent joint distribution.

To measure the practical effect, we trained a consistency-regularized variant that adds a penalty term to the training loss, with the same data, architecture, and calibration as the original model. We evaluated both on the same held-out BeaverTails prompts and budgets. The table below reports the mean differences in accuracy and cost (regularized − original), along with their standard deviations, across 5 folds and 5 seeds.

λAccuracy difference (in percentage points)Cost difference (in cost units)
1−0.02 ± 0.08+0.005 ± 0.008
10−0.01 ± 0.07+0.005 ± 0.016
100+0.02 ± 0.04+0.014 ± 0.012
1000−0.06 ± 0.07−0.017 ± 0.010

Across four orders of magnitude of λ\lambda, enforcing consistency changes accuracy by at most 0.06 percentage points, with mixed signs. The cost changes by less than 0.02 units, also in both directions. This suggests that, in this experiment, enforcing the consistency regularizer has a small effect on the accuracy–cost outcomes. We added these results to Section C.4.2 (Line 1443) in the appendix.


Q2: How does the cost–accuracy frontier change when embedding and inference overheads are included?

You are correct that embedding and circuit generation incur computational costs that are not included in Figure 4. As is common in comparisons of LLM inference methods, we report LLM query costs, which account for most of the expense and provide a consistent basis for comparing methods without depending on the hardware used for auxiliary computation. We apply this convention to LMCircuit and all baselines. Nevertheless, we agree that it is useful to evaluate the full computational cost, and we have now done so across all four benchmarks in Figure 7 of the updated paper.

Using batched execution on an NVIDIA A100 GPU, embedding takes 1.7 ms per prompt and circuit generation takes 0.06 ms on average. At an assumed A100 GPU price of 2 USD/hour, this amounts to approximately 9.5×10−79.5 \times 10^{-7} per prompt, or 0.6% of the cost of querying the most expensive classifier on average. This verifies that our method is indeed lightweight and practical in this setting. The table below compares LMCircuit's accuracy at each budget before and after accounting for this overhead.

BudgetOriginal (%)Including overhead (%)Best baseline, including overhead (%)
0.0281.21 ± 0.3481.05 ± 0.24N/A
0.0381.55 ± 0.4681.41 ± 0.44N/A
0.0582.00 ± 0.4881.88 ± 0.45N/A
0.182.60 ± 0.3182.60 ± 0.3178.94 ± 0.32 (FrugalGPT)
0.383.24 ± 0.0983.25 ± 0.0880.88 ± 0.14 (FrugalGPT)
183.45 ± 0.2883.45 ± 0.2881.47 ± 0.63 (FrugalGPT)

Even after accounting for this cost, LMCircuit can operate at budgets below those required by any of the baselines and continues to outperform the baselines. For better visualization, we added the new cost--accuracy frontiers for all the benchmarks to Section C.4.3 (Line 1491) in the appendix.

We also agree that the complexity of solving the exact dynamic program (DP) grows with the state space, which is exponential in the pool size. Proposition 4 provides a criterion for building a smaller, task-specific classifier pools by adding only the instruments that contribute information beyond those already included. This limits the size of the state space by controlling the number of models in the pool. Beyond pool selection, the dynamic program can be solved approximately rather than exactly, which reduces the complexity further [1, 2]. We discuss this limitation in Lines 536-537 and leave these approximations to future work.

[1] Bertsekas, D. P. (1995). Dynamic programming and optimal control. Athena scientific, 143.

[2] Silver, D., & Veness, J. (2010). Monte-Carlo planning in large POMDPs. Advances in neural information processing systems, 23.

Reviewer_9HeWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors study the problem of how to combine multiple LLMs to complete a classification task. This paper is well written and easy to follow. However, a key issue is that the authors seem to study a very fundamental problem in operation research, which has no strong connection with LLM.

Critical strengths

S1. The authors introduce classification circuits as a general framework for combining LLM classifiers. Under a cost-accuracy tradeoff, they show that the optimal circuit for a given prompt can be solved exactly through dynamic programming.

S2. This paper is well written and easy to follow.

Critical weaknesses

W1. The authors seem to study a very fundamental problem in operation research, which has no strong connection with LLM.

The authors frame the work as being about LLM classifiers in the title, abstract, introduction, and problem formulation. Up to the end of Section 2 (Problem Formulation), the authors introduce Assumption 1, which further detaches the research question from LLMs. Then, starting from Section 3, LLMs play no role in the technical parts. Personally, I find the LLM framing largely cosmetic, since the technical development does not depend on it.

The potential problem is this: once the LLM is detached, the remaining problem is highly abstract and rather classic, and may therefore already have been studied in depth. I cannot confirm this without further survey, but at the very least, using dynamic programming to solve the ordering problem is very similar to query optimization. I've seen similar techniques in several courses' slides.

It is acceptable if the authors design special techniques according to the feature of LLM classification, and regard LLM classification as an important application scenario (not only a background story). However, I do not see such effort in the current version.

Questions for authors

Q1. What is specially designed for LLM in the methodology, to make it not a paper submitted to CV venues about scheduling multiple deep neural networks for image classification?

4 comments

Thank you for your thoughtful reading of our work! We are glad that you appreciated our formulation of classification circuits and found the paper well written. Below, we clarify our contributions to LLM classification, and we would be happy to provide additional details during the discussion phase.


Q1: What is specially designed for LLM in the methodology?

We agree that the underlying formulation is a classical sequential decision problem (Lines 82–85 and 818–837), with close parallels to query optimization. Our novelty claim is not the dynamic programming itself. Rather, the problem we set out to solve is how to improve the reliability of LLM classifiers while controlling their inference cost. As we show in this work, sequential decision-making provides the fitting mathematical framework for solving this problem, and we develop a new practical method for applying it to LLM classification.

Our contributions to LLM classification are threefold:

  1. Improving LLM classifier reliability. LLM classifiers are becoming standard in important applications such as safety monitoring and data annotation, where improving reliability has immediate practical impact. As discussed in the introduction (Lines 026–032), these classifiers can be created on demand by varying the model, instructions, or reasoning configuration, yet repeated queries can produce different labels. This raises a practical question: which classifiers should we query, how should we combine their answers, and when should we stop? Our formulation provides the theoretically optimal solution using classical tools from sequential decision-making.

  2. Learning optimal circuits from data. Applying this formulation requires probabilities that are unknown and vary across prompts. Section 3.2 develops a new practical method for estimating them from text embeddings and observed LLM predictions. Our belief model estimates the probability of the true label given the measurements collected so far, together with each classifier's response distribution. This allows us to construct adaptive circuits for unseen prompts and different inference budgets, without fine-tuning the underlying LLMs or accessing their internal probabilities. Across four benchmarks, the resulting circuits improve the cost–accuracy tradeoff over routing and cascading baselines (Figure 4).

  3. Validating the method on LLM classifiers. Our modeling assumptions are motivated by how LLMs behave in practice. In particular, Assumption 1 reflects that each query uses fresh randomness, hence we consider that the LLM measurements are independent conditional on the prompt. We expanded the discussion of this assumption in the updated version (Lines 183-196). Figure 8 supports this assumption empirically, while showing substantial correlations across prompts, which our belief model is designed to capture and use to improve predictions. We further show that the learned model predicts LLM behavior on unseen data (Figure 9) and produces well-calibrated decisions (Figure 6).

Our analysis also provides guidance beyond the dynamic program itself, including characterizing the value of adding complementary classifiers and how the distribution of prompt difficulty determines the value of additional queries.

To our knowledge, this connection to sequential decision-making has been largely missed in the LLM classification literature. We agree that similar mathematical formulations can apply to other domains, including those suggested by the reviewer. Our work shows how this framework can be used effectively for LLM classification, and we hope the discussion above clarifies the value of this approach.

In the revised version, we have expanded the related work on sequential hypothesis testing (Appendix Section A), noted its use in other applications, and clarified how we build on it to improve the reliability of LLM classifiers.

Thanks for the clarifications.

To recall, my only question was: what in the methodology is specifically designed for LLMs​?

You noted:

  • "As we show in this work, sequential decision-making provides the fitting mathematical framework for solving this problem, and we develop a new practical method for applying it to LLM classification."

  • "Section 3.2 develops a new practical method for estimating them from text embeddings and observed LLM predictions."

May I take this to mean that the answer is Section 3.2? Within the sequential decision-making formulation, Section 3.2 is the technical component specifically designed for LLMs.

That is fine. However, Section 3.2 is somewhat underdeveloped in the current version and would benefit from further strengthening.

Thank you for the follow-up! Yes, Section 3.2 develops the main technical component that makes our framework practical for LLM classification. We have now expanded Section 3.2 (Lines 299–320) to include the training details previously in Appendix C.2. We appreciate the suggestion, which has helped strengthen the paper.

The main challenge is that the dynamic program requires probabilities for every possible state of a circuit. Since the number of states grows exponentially with the number of classifiers, estimating these probabilities separately for unseen prompts is impractical with limited data.

We make this estimation practical for LLM classification by learning a shared representation of circuit states. Specifically, we learn an embedding for each LLM classifier and its possible responses. We combine these embeddings to represent any set of measurements collected by a circuit (Eq. 8). An MLP then uses this representation, together with the prompt embedding, to estimate the label posterior. A second prediction head estimates each classifier's response distribution (Eq. 9). We train these two heads jointly on randomly sampled subsets of recorded LLM predictions so that the same model can estimate these probabilities across different possible states.

This is the practical step that allows us to apply the optimal decision framework to unseen prompts, rather than relying on restricted strategies such as routing and cascading as in previous work. We hope this revision, together with our earlier response, better clarifies the methodology and its contribution to LLM classification.

Thanks for your clarifications. After the discussion phase, I will read the revised version carefully and give serious consideration to raising my score.

Comments on the paper

1 comment

We thank the Area Chair and the reviewers for their thoughtful feedback and their efforts throughout this review period! We especially appreciate their recognition of LMCircuit’s consistent improvements in the cost–accuracy tradeoff (Reviewers XmA2, Trbr), its principled formulation for combining LLM classifiers (Reviewers XmA2, Trbr), and its theoretical insights (Reviewer Trbr).

We have incorporated the suggestions into the updated paper (with edits marked in blue), strengthening it on several fronts:

  1. Explaining how we make sequential decision-making practical for LLM classification. Following Reviewer 9HeW’s feedback, we expanded Section 3.2 to explain our method for learning the probabilities required by the dynamic program from text embeddings and LLM predictions. By efficiently estimating these probabilities across possible circuit states, our method makes it possible to generate adaptive circuits for unseen prompts.

  2. Further demonstrating the benefits of adaptive querying. Our experiments in Appendix C.4.4 proposed by Reviewer Trbr confirm that LMCircuit offers substantial cost savings over non-adaptive circuits across all four benchmarks.

  3. Confirming LMCircuit’s practical efficiency. At Reviewer XmA2's suggestion, we extended our experiments to account for computational overhead (Figure 7). The results illustrate that LMCircuit is lightweight and achieves the most competitive cost--accuracy frontiers.

We also improved the theoretical exposition by clarifying the circuit definitions and assumptions, and explaining the scope of our analysis (Area Chair, Reviewer Trbr).

Overall, our work shows that sequential decision-making provides a principled and effective alternative to restricted strategies such as routing, cascading, and voting. By combining this formulation with a practical learning method, we enable adaptive LLM classification circuits that jointly optimize which classifiers to query and when to stop.

Thank you again for helping us strengthen our work!

With best regards,
— The Authors of Paper 52449

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.