Hierarchical Group-Conditional Conformal Risk Control for Selective Prediction in Language Models
HG-CRC controls hierarchical group risk; on ARC Challenge it achieved 0% empirical violations and WGER=0.
Key Findings
Methodology
HG-CRC calibrates one threshold for every node in a user-defined group tree. It uses the Clopper–Pearson upper confidence bound for selective risk and assigns each node confidence δ/|H|, enabling simultaneous control through Bonferroni correction. At inference, a leaf-first policy selects the deepest certified ancestor whose score threshold accepts the example; otherwise it falls back toward the root. The paper studies both data-efficient in-sample residual calibration and formally valid split calibration.
Key Results
- On ARC Challenge, Qwen3-4B and Llama-3.1-8B-Instruct achieved a 0% empirical violation rate and WGER=0. Because the evaluation used 500 bootstrap trials, these zeros are only empirical upper bounds: a true violation rate of approximately 0.6% remains compatible with observing no violations.
- Under mild group-composition shift, global CRC violated the risk budget in up to 47% of trials. HG-CRC reduced this failure, but participation fell by 22–37 percentage points relative to global CRC. On MMLU-Pro, some models abstained entirely, while Llama retained WGER=0.014.
- Ablation showed that hierarchy depth matters: removing the difficulty level returned ARC violations to approximately 11%. Bonferroni is required for the theoretical guarantee; its empirical cost was small for ARC’s five-node hierarchy but became material with many nodes.
Significance
The work shifts LLM reliability from average population performance toward protection of every operational subgroup. This matters in medical question answering, educational assessment, and legal retrieval, where difficult topics or minority domains can be hidden by aggregate accuracy. HG-CRC offers a post-hoc safety layer requiring only a labeled calibration set, not retraining. Its central practical message is also cautionary: stronger groupwise safety is purchased with lower availability, and the cost depends on benchmark difficulty, score calibration, and group sample size.
Technical Contribution
The paper extends CRC from one marginal constraint to simultaneous, hierarchical conditional-risk constraints. Its residual membership map ρv removes examples already answered by certified descendants, matching calibration with the distribution actually deployed at node v. Proposition 1 gives an unconditional simultaneous guarantee for depth-split calibration; Proposition 2 gives exact validity for test-point-augmented in-sample calibration; Lemma 1 bounds the deployed plain in-sample procedure by an additive εn slack. This explicitly separates formal validity from statistical efficiency.
Novelty
Relative to Mondrian conformal prediction and flat groupwise CRC, HG-CRC is, to the authors’ knowledge, the first framework for hierarchical group-conditional selective risk control in LLMs. Its fundamental innovation is not merely estimating one threshold per group, but combining residual calibration, Bonferroni simultaneous inference, and leaf-first routing. The parent node therefore handles only unresolved residual cases rather than competing with every specialized child threshold.
Limitations
- Performance is benchmark-dependent. On MMLU-Pro, Qwen and Gemma may abstain entirely, while Llama has WGER=0.014, indicating that valid risk control can collapse into very low coverage when uncertainty scores are poorly aligned with task correctness.
- The deployed in-sample algorithm is not unconditionally distribution-free because descendant thresholds and residual membership are estimated from the same data. Its guarantee depends on threshold regularity and measured εn; sparse Gemma leaves can have weaker conditional slack.
- Nodes require at least 30 calibration examples, while deeper trees and Bonferroni correction reduce statistical power and participation.
Future Work
Future research should develop adaptive hierarchies, less conservative simultaneous testing, and online recalibration under domain or prompt shift. It should also assess automatically inferred difficulty labels, overlapping and dynamically changing groups, multilingual settings, and open-ended generation. A key engineering goal is to optimize the safety–coverage frontier so that conformal guarantees do not produce systems that are technically safe only because they rarely answer.
AI Executive Summary
Large language models are deployed across heterogeneous populations, yet standard Conformal Risk Control (CRC) usually protects only the aggregate answering population. A model can meet a global error budget while systematically overexposing difficult questions, particular domains, or minority linguistic styles. The paper demonstrates this structural gap: under mild shifts in group composition, global CRC violated the per-group budget in up to 47% of bootstrap trials.
The authors introduce Hierarchical Group-Conditional CRC (HG-CRC), a post-hoc calibration framework built on a user-defined tree of groups. Each node receives a threshold calibrated with a Clopper–Pearson upper bound, while Bonferroni correction assigns confidence δ/|H| so all nodes are protected simultaneously. At deployment, a leaf-first policy uses the deepest certified threshold that accepts an example and falls back to coarser ancestors otherwise. Residual calibration removes cases already answered by certified descendants, aligning each threshold with its actual deployment population. No retraining is required.
Across Qwen3-4B, Llama-3.1-8B-Instruct, Gemma-3-4B, ARC Challenge, and MMLU-Pro, HG-CRC produced 0% empirical violations and WGER=0 for Qwen and Llama on ARC. With 500 bootstrap trials, however, the implied empirical upper bound is about 0.6%, not a certified zero. Removing the difficulty level restored violations to roughly 11%, while participation cost 22–37 percentage points versus global CRC. On MMLU-Pro, some systems abstained entirely and Llama retained WGER=0.014. The study therefore establishes a credible route to groupwise safety, while making the availability cost and benchmark dependence impossible to ignore.
Deep Analysis
Background
Selective prediction lets a model answer only when sufficiently confident, trading participation ϕ for lower selective risk R. Angelopoulos et al. (2023) showed that CRC can use held-out data and Clopper–Pearson bounds to control marginal risk under exchangeability. Mondrian conformal prediction and related fairness work address subgroup structure, but mostly for prediction sets or independent flat groups. The unresolved issue is selective LLM answering under heterogeneous and shifting group composition.
Core Problem
Given classifier f, uncertainty score s(x), risk target α, and confidence 1−δ, the goal is to control risk not only globally but for every node v in a hierarchy. This is difficult because test group proportions may shift, descendants preferentially answer easy examples, and a parent’s deployment residual becomes harder than its calibration population. Testing many nodes independently also inflates the family-wise error rate.
Innovation
HG-CRC contributes four linked ideas. First, it calibrates thresholds separately across a hierarchy rather than using one global threshold. Second, Bonferroni correction converts nodewise CRC into a simultaneous guarantee. Third, leaf-first routing ensures specialized groups receive priority. Fourth, residual calibration removes examples claimed by certified descendants, preventing parent thresholds from being calibrated on an easier mixture than the one they will actually serve. These mechanisms jointly target both fairness and distributional mismatch.
Methodology
- �� Inputs: model f, score s, calibration set Dcal, hierarchy H, α, δ, and minimum node size Nmin=30.
- �� Node calibration: for candidate threshold τ, count errors kv(τ) and answered cases nv(τ), then compute Uv(τ)=Beta⁻¹(1−δ/|H|;kv+1,nv−kv). Select the largest certified threshold among 100 calibration-score quantiles.
- �� Residuals: define ρv(x)=mv(x)∏c(1−mc(x)gτc(x)); parent v retains only examples not answered by certified descendants.
- �� Prediction: scan the ancestry path from leaf to root; answer with the first certified node satisfying s(x)≤τ*v, otherwise abstain.
- �� Guarantees: depth-fold split calibration yields Proposition 1’s unconditional simultaneous bound; plain in-sample calibration is covered by Lemma 1 with εn slack.
Experiments
The study evaluates Qwen3-4B, Llama-3.1-8B-Instruct, and Gemma-3-4B on ARC Challenge and MMLU-Pro. Eight configurations probe IID generalization, group heterogeneity, mixture, domain, prompt, and difficulty shift, label noise, and quantization. Canonical targets are α=0.1 for ARC and α=0.3 for MMLU-Pro. Baselines include global CRC and flat groupwise calibration. Metrics are violation rate, WGER, and participation; ablations vary hierarchy depth, Bonferroni correction, uncertainty score, and split versus in-sample calibration.
Results
Global CRC reached as many as 47% violating trials under group-composition shift, whereas HG-CRC gave 0% empirical violations and WGER=0 for Qwen and Llama on ARC. Removing difficulty increased violations to about 11%. Participation was 22–37 percentage points lower than global CRC, and split calibration added roughly 0.6–5.5 points relative to in-sample calibration. MMLU-Pro was substantially harsher: some models abstained completely, while Llama retained WGER=0.014. Gemma’s poor score calibration caused conservative abstention rather than uncontrolled errors.
Applications
HG-CRC can govern educational assessment assistants by subject, grade, and difficulty, routing uncertain cases to teachers. In medical or legal retrieval, nodes can represent specialty, case type, and complexity, with unserved cases escalated to experts. Deployment requires labeled, reasonably representative calibration data, reliable group metadata, and an uncertainty score correlated with correctness. The method is a model-agnostic post-processing layer, not a replacement for capability training or monitoring.
Limitations & Outlook
The formal results depend on exchangeability, meaningful group definitions, and sufficient calibration data. Continuous drift, overlapping groups, or biased difficulty labels may invalidate practical alignment. Bonferroni becomes conservative as the hierarchy expands, and split calibration sacrifices data at each depth. MMLU-Pro shows that safety may manifest as near-total abstention; sparse residual cells also weaken εn-based guarantees. Future work should improve multiple testing, online recalibration, and evaluation on open-ended generation.
Plain Language Accessible to non-experts
Imagine a large school where teachers may grade a student’s answer or say, “I’m not sure.” The school sets a rule: among the answers teachers choose to grade, fewer than 10% should be wrong. Ordinary CRC checks the average across the whole school. If the easy class performs brilliantly while the difficult class suffers, the school average can still look safe even though one class is being treated unfairly.
HG-CRC builds a management tree: school, grade, class, and perhaps difficulty level. Each level gets its own confidence rule, and all rules are checked together. When a new answer arrives, the system first uses the most specific reliable class rule. If that rule refuses or has not been certified, it moves upward to the grade and school rules.
The clever part is that the larger group is judged only on questions left over after smaller groups have acted. It does not pretend that the remaining questions are an average mixture. This protects subgroups more honestly, but teachers will say “I’m not sure” more often. On ARC Challenge, Qwen and Llama achieved zero observed violations, yet the price was lower availability and weaker behavior on MMLU-Pro.
ELI14 Explained like you're 14
Think of an AI as a gamer answering quiz questions. It can answer or press “skip.” If it answers every question, it may look impressive but confidently make lots of mistakes. A normal safety check looks at the average score. That can hide a problem: easy questions are fine, while players asking hard science questions keep getting tricked.
HG-CRC is like giving the game several shields. One shield is for a topic, another for difficulty, and another for the whole quiz. The most specific shield gets the first chance. If it is not trustworthy, the system tries a bigger shield. If no shield accepts the question, it skips it. That sounds annoying—but isn’t skipping better than confidently giving a wrong medical or legal answer?
The system also keeps track of which questions smaller shields already handled. Bigger shields judge only the leftovers, instead of mixing easy and hard questions together. In experiments, Qwen3-4B and Llama-3.1-8B-Instruct had 0% observed ARC violations and WGER=0.
But there is no free power-up! Compared with global CRC, participation dropped by 22–37 percentage points. On MMLU-Pro, some models almost never answered. So the real challenge is balancing safety and usefulness: an AI that never answers is safe in one sense, but not very helpful. The goal is a smart “skip,” not a permanent mute button!
Glossary
Conformal Risk Control (CRC)
A calibration method that uses held-out data to control prediction risk without specifying a parametric model distribution. Its standard guarantee is marginal and relies on exchangeability.
HG-CRC extends CRC from one global threshold to simultaneous node-specific thresholds.
Selective prediction
A prediction system may answer some inputs and abstain on others. Its key trade-off is between participation and conditional error risk.
Abstention is HG-CRC’s mechanism for satisfying groupwise budgets.
Clopper–Pearson upper bound
An exact one-sided confidence bound for a binomial error rate, expressed through a Beta quantile. It is conservative but distribution-free for the calibration count.
A node is certified only when its bound is at most α.
Bonferroni correction
A family-wise error control method that assigns δ/|H| to each of |H| tests. It guarantees that all nodewise claims hold jointly with probability at least 1−δ.
HG-CRC uses it for simultaneous hierarchical risk control.
WGER
Worst-Group Excess Risk: max_v max(R_v−α,0). WGER=0 means no evaluated hierarchy node exceeds the target risk.
The paper uses WGER as its principal group-safety summary.
Residual calibration
Calibration of a parent node only on examples not answered by certified descendants. It aligns the calibration population with the parent’s actual deployment residual.
Residual calibration is central to HG-CRC’s leaf-first routing.
Open Questions Unanswered questions from this research
- 1 How can one retain simultaneous guarantees when there are many nodes, sparse cells, overlapping groups, and continuous drift without the severe conservatism of Bonferroni? More powerful hierarchical multiple-testing methods are needed.
- 2 MMLU-Pro’s low coverage may reflect model capability, score miscalibration, or hierarchy design. Separating these causes requires broader multilingual, cross-task, and online-shift evaluations.
- 3 Automatically inferred difficulty labels may themselves be biased and could redirect risk to poorly represented groups. Reliable, auditable hierarchy construction remains unresolved.
Applications
Immediate Applications
Educational assessment assistant
Schools can build nodes by subject, grade, and difficulty, then calibrate answer thresholds on labeled exam data. The assistant grades only when the relevant node is certified and otherwise routes the item to a teacher, preventing easy-question averages from hiding systematic failures on difficult content.
Medical and legal triage
Hospitals and law firms can define nodes by specialty, case type, and complexity. A calibrated model answers routine cases while uncertain or uncertified residual cases are escalated to professionals. Prerequisites are representative labels, stable metadata, and a score correlated with correctness.
Long-term Vision
Auditable group-safe model gateways
Future inference gateways could monitor WGER by subgroup and trigger targeted recalibration after domain or prompt shifts. Combined with more efficient simultaneous inference, this could make reliability and fairness operationally auditable without forcing every safe system into near-total abstention.
Abstract
Large language models serve heterogeneous populations structured by domain, topic difficulty, and linguistic style. Conformal risk control (CRC) gives rigorous marginal risk guarantees for selective prediction with abstention, but marginal guarantees do not imply per-group ones: a model can meet the population budget while systematically over-exposing subgroups to errors. Under mild shift in group composition, standard CRC violates the budget in up to 47% of trials. We propose HG-CRC (Hierarchical Group-Conditional CRC), a post-hoc calibration framework enforcing simultaneous risk guarantees across all nodes of a user-defined group hierarchy. It applies a Bonferroni correction over nodes and a leaf-first policy that uses the most specific applicable threshold, falling back to coarser nodes when a finer one is uncertified or rejects the example. It needs only a held-out calibration set, with no retraining. We evaluate on three models (Qwen3-4B, Llama-3.1-8B-Instruct, Gemma-3-4B) and two benchmarks (ARC Challenge, MMLU-Pro) across eight configurations probing IID generalization, heterogeneity, mixture/domain/prompt/difficulty shift, label noise, and quantization. Main result: HG-CRC reaches an empirical 0% violation rate and WGER=0 on ARC Challenge for high-accuracy models (Qwen3-4B, Llama-3.1-8B). At 500 bootstrap trials these zeros are empirical upper bounds (true rate up to 0.6%), not certified. Results are benchmark-specific: on MMLU-Pro these models abstain entirely or (Llama) retain WGER=0.014. Gemma-3-4B, poorly calibrated here, degrades gracefully by abstaining. Participation cost vs. global CRC is 22 to 37 points. Ablations show hierarchical depth clears the budget: removing difficulty level returns violations to about 11%. Bonferroni is needed for the theoretical guarantee, though its empirical effect matters only with many nodes.