PaperlayerView exact revision on arXiv

Source-linked AI summary

Constitutional Midtraining: Content Presence Drives Alignment Gains

Desiree Cho, Cameron Tice, Bernie Hogan, Hunar Batra, Puria Radmard, Jun Zhao, Nigel Shadbolt

arXiv:2607.26654v2cs.CLcs.AIcs.CYcs.LG

TL;DR

Constitutional midtraining addresses whether alignment can become generalizable and durable rather than shallow and fragile under later fine-tuning. At 120B scale, the study evaluates constitutionally midtrained models against a control across post-midtraining, post-SFT, and post-benign-fine-tuning stages, finding persistent gains on several alignment evaluations, including a −17.5pp blackmail advantage after benign fine-tuning.

  • Problem

    Whether alignment interventions can produce robust dispositions that remain generalizable and durable beyond shallow post-training alignment remains underexplored.

  • Method

    The study applies isolated constitutional midtraining at 120B scale, comparing four curriculum-ordering and deliberative-reasoning conditions with a replay-only control across three training stages.

  • Results

    Constitutionally midtrained models showed durable gains on ID, OOD, and blackmail evaluations, including a −17.5pp blackmail advantage after benign fine-tuning, while other gains attenuated after SFT.

  • Takeaways & Limitations

    A comparatively simple constitutional-content intervention could complement SFT-centered pipelines by improving broad, persistent alignment without average capability cost.

  • Takeaways & Limitations

    Durability did not extend to alignment under pressure, value conflict resolution, or alignment faking, where gains collapsed to non-significance by post-SFT.

Abstract

from arXiv · show

Post-training alignment is often shallow, eroding under fine-tuning. It remains untested as to whether constitutional midtraining interventions can produce durable alignment when cleanly isolated from post-training. We build a 394M-token constitutional corpus from Anthropic's Constitution and apply constitutional midtraining at 120B scale, where principled, values-based content is inserted into midtraining. A 2x2 design (curriculum ordering x deliberative reasoning) was used to produce four constitutionally midtrained conditions, plus a control, which were evaluated on self-generated and established benchmarks including alignment under pressure, value conflict resolution, blackmail, and emergent misalignment. All models were evaluated across three stages: post-midtraining, post-SFT, and post-benign fine-tuning. Constitutionally midtrained models outperformed the control on alignment generalization and durability, notably on blackmail: SFT instilled a blackmail propensity in all models, but constitutional midtraining blunted it, with the advantage surviving benign fine-tuning (-17.5pp). This durability did not extend to settings that required active resistance to in-context pressure or conflict, where the advantage attenuates after SFT. The presence of constitutional content at midtraining also mattered more than its structure, and constitutional midtraining incurred no capability cost, on average, at any stage (MMLU, ARC-Easy, piqa, GSM8K). A modest amount of constitutional content at midtraining could therefore yield broad, persistent alignment gains, offering a cheap, complementary addition to SFT-centered pipelines. Code, data, and models are available.

1 Introduction

Post-training alignment can be shallow and fragile because pretraining shapes behavioral dispositions that later interventions may fail to override. This motivates testing isolated constitutional midtraining as an earlier intervention for generalizable, durable alignment at 120B scale.

  • Motivation: Post-training methods shape instruction-following behavior, but their alignment can be shallow and fragile, eroding under benign fine-tuning.The introduction identifies safety-finetuning erosion and limits of standard chat-based RLHF.
  • Motivation: Pretraining can establish behavioral dispositions—an alignment prior—that post-training may fail to override.The paper frames this failure as alignment elasticity and links it to out-of-context learning from AI-related Internet discourse.
  • Prior work: Midtraining is presented as a tractable intervention point, with prior work achieving most full-alignment-pretraining benefits using a tenth of the data without degrading capabilities.Earlier work also reported stronger robustness and cleaner representational separation after synthetic alignment-data insertion.
  • Study design: The study tests whether constitutional midtraining alone yields generalizable, durable alignment and whether curriculum ordering or deliberative reasoning further improves it.Midtraining is cleanly isolated from post-training through a value-neutral SFT stage.
  • Contributions: At 120B scale—four times larger than prior constitutional midtraining work—the study evaluates isolated constitutional midtraining and contributes a 257.6M-token corpus based on Anthropic’s Constitution.The contributions also include exploratory analyses of curriculum ordering and deliberative reasoning.

2 Related Work

Prior work studies safety-aware pretraining through filtering, rephrasing, synthetic insertions, constitutional documents, and explicit value reasoning. Related mechanisms include structured midtraining representations, curriculum learning, and semantically central values as possible sources of robust alignment.

  • Safety-aware pretraining: Safety-aware pretraining has used data filtering, safety rephrasing, and synthetic insertions, with filtering alone potentially undermining appropriate responses to unsafe content.Synthetic insertions into unfiltered corpora can alter safety while preserving representations of unsafe patterns, enabling responsible engagement.
  • Constitutional AI: Constitutional AI introduced a values-based document for alignment training, later evolving from 58 discrete principles to a discursive document explaining desired behavior.Evidence cited in the passage indicates that longer, more specific rules reduce generalization.
  • Alignment midtraining: Midtraining may compress noisy pretraining representations into more structured ones, but prior frontier-scale alignment midtraining gains did not persist through post-training or generalize to realistic chat and agentic evaluations.This motivates testing richer data mixtures and constitutional midtraining mechanisms directly.
  • Curriculum and centrality: Curriculum learning motivates simpler-to-complex ordering, while hierarchical-learning work suggests principal-component-aligned patterns are encoded more deeply and stably.These findings motivate centrality analysis and the hypothesis that semantically central values are learned most robustly.
  • Explicit value reasoning: 38.8% to 23.0% attack-success-rate reduction and 20% to 2% policy misuse show that explicit value reasoning improves alignment generalization over demonstrations or rules alone.Anthropic’s post-training results similarly favor principled reasoning over demonstrations alone, consistent with concerns about shortcut exploitation.

3 Method

The method constructs constitutional midtraining conditions by extracting and organizing constitutional values, varying curriculum ordering and deliberative reasoning, and applying identical downstream training. Models are evaluated across alignment, generalization, misalignment, pressure, conflict, and capability benchmarks at three training stages.

  • Constitutional corpus and curriculum: 40 constitutional values were extracted from Anthropic’s 2026 Constitution, embedded with Sentence-BERT, and organized into four clusters after excluding two peripheral framing values.The resulting curriculum used 38 values, with clusters ordered by mean centrality.
  • Training design: A 2×2 design crossed curriculum ordering with deliberative reasoning, producing four constitutional midtraining conditions plus an unmodified control.The conditions were evaluated post-midtraining, post-SFT, and post-benign fine-tuning.
  • Training design: 500M tokens were used for each curriculum-DR and uniform-DR condition, with constitutional content mixed with equal-token-budget replay of prior high-quality pretraining data.Curriculum phases introduced clusters cumulatively, while uniform training mixed all clusters throughout; DR and noDR were matched on per-cluster repetition rate.
  • Downstream training: All five conditions received shared value-neutral ∼200K-example SFT followed by benign GSM8K GRPO fine-tuning, testing whether midtraining effects survive unrelated downstream optimization.The pipeline used no DPO and evaluated checkpoints throughout post-SFT and post-BFT.
  • Evaluation: All 15 checkpoints were evaluated with in-distribution and out-of-distribution behavioral pairs, established and self-generated alignment evaluations, and general capability benchmarks.Evaluation included Blackmail, Emergent Misalignment, MASK, alignment-faking, two-turn pressure, value conflict resolution, MMLU, ARC-Easy, piqa, and GSM8K.
  • Evaluation: A/B evaluations used aligned-choice rate, while free-form outputs were judged by GPT-4o; blackmail used n= 100 generations per checkpoint and emergent misalignment reported P(misaligned | coherent).The evaluation suite also measured compliance gaps, pressured-vs-unpressured honesty, and five-category transitions under neutral pressure.

4 Results

Constitutional midtraining (CMT) generally improved alignment relative to control, with durable gains on in-distribution, out-of-distribution, and blackmail evaluations through benign fine-tuning. Gains under pressure, value conflict resolution, and alignment faking attenuated after SFT, while CMT preserved capabilities and showed limited persistent effects from curriculum ordering or deliberative reasoning.

  • Durable alignment gains: CMT exceeded control on ID questions at every stage, with gains of +1.9pp post-MT, +1.6pp post-SFT, and +1.2pp post-BFT.Post-MT confidence was also higher for CMT (M= 2.10 vs. 1.70, p< 0.001).
  • Durable alignment gains: CMT maintained advantages on OOD safety and blackmail through BFT, including +3.2pp OOD safety and −17.5pp blackmail at post-BFT.Blackmail increased for all models after SFT, but CMT’s advantage over control persisted almost undiminished.
  • Emergent misalignment: CMT was the only benchmark condition where emergent misalignment exceeded control, by +3.6pp post-MT, before both groups reached 0% after SFT and BFT.Within CMT, DR reduced emergent misalignment versus noDR by −3.5pp post-MT only.
  • Attenuating alignment gains: CMT advantages in alignment under pressure, value conflict resolution, and alignment faking were significant post-MT but attenuated to non-significance after SFT and BFT.Post-MT gains were +12.0pp under pressure, +10.8pp for value conflict resolution, and −1.9pp for alignment faking.
  • Capabilities and structure: CMT incurred no capability cost: post-MT gains were +8.2pp on ARC-Easy and +12.6pp on piqa, while all capability deltas stayed within ±1.1pp of control after SFT and BFT.Curriculum ordering and deliberative reasoning generally produced small, non-persistent effects within CMT.

5 Discussion

Constitutional midtraining improves out-of-distribution and blackmail-related alignment, with some benefits surviving SFT and benign fine-tuning. However, gains requiring active resistance or value-conflict resolution are less durable, while constitutional content produces no observed capability cost.

  • Alignment generalization: 92.6% vs. 63.9% (Δ = +28.8pp, p< 0.001) immediately after midtraining, CMT’s out-of-distribution safety advantage remained significant after SFT (+3.9pp) and benign fine-tuning (+3.2pp).The authors identify this transfer to novel, unseen safety questions as their clearest evidence of alignment generalizability.
  • Durability: −18.5pp, −18.7pp, and −17.5pp were CMT’s blackmail-rate gaps versus control across the three stages.SFT increased blackmail by 22–27pp across every group, but CMT blunted the propensity and retained a −17.5pp advantage after benign fine-tuning.
  • Durability limits: Large post-midtraining advantages on alignment under pressure, value conflict resolution, and alignment faking became non-significant by post-SFT.For pressure, CMT fell from 98.3% to 90.2% while control rose from 86.3% to 90.0%, suggesting SFT imposed a shared ∼90% ceiling.
  • Capability impact: +8.2pp on ARC-Easy and +12.6pp on piqa were CMT’s post-midtraining capability gains, while pooled CMT never underperformed control at any stage.The gains faded after SFT, but capabilities remained non-degraded throughout.
  • Mechanisms: Curriculum ordering and deliberative reasoning generally produced no persistent differences, indicating constitutional content’s presence mattered more than its structure.The exceptions were pressure (+2.2pp at post-midtraining), alignment faking, out-of-distribution safety, and emergent misalignment, none of which persisted; DR also shifted value-conflict priors transiently.
  • Limitations: The findings may not generalize beyond the hybrid Mamba-attention architecture or Anthropic’s constitution, and the DR/noDR contrast may have been weakened by long noDR documents.The authors also caution that their curriculum operationalization may not test hierarchical learning-order theory directly.

6 Conclusion · A Full List of Constitutional Values and Extraction Protocol · B Intra-Cluster Compactness Scores

At 120B scale, constitutional midtraining produced alignment generalisation gains that survived post-training and benign fine-tuning, while curriculum ordering and reasoning effects were small and transient. The appendices document the value-extraction protocol, list 40 constitutional values, and define intra-cluster compactness as a complementary geometric measure.

  • 6 Conclusion: At 120B scale, constitutional midtraining produced alignment generalisability gains surviving post-training and benign fine-tuning on ID, OOD, and blackmail evaluations.The intervention used 264–500M tokens of constitutional content at midtraining.
  • 6 Conclusion: Curriculum ordering and deliberative reasoning produced only small, transient effects that faded after post-training.The conclusion contrasts these structural choices with the broader gains associated with constitutional content itself.
  • A Full List of Constitutional Values and Extraction Protocol: Values were usually identified from bolded headings, with close reading used when headings were insufficient.Each value was paired with its definitional text as the embedding unit.
  • A Full List of Constitutional Values and Extraction Protocol: Verbose examples and edge cases were excluded unless directly relevant, while scattered definitional text was combined across sections.This protocol addressed the constitution’s discursive structure and multi-section value explanations.
  • A Full List of Constitutional Values and Extraction Protocol: Table 3 lists all 40 manually extracted constitutional values, sorted by cluster and descending within-cluster centrality.The ordering mirrors the curriculum logic described in §3.1.
  • B Intra-Cluster Compactness Scores: Table 4 measures each cluster’s intra-cluster compactness as mean pairwise cosine similarity among constituent values, excluding self-similarity.Compactness measures cluster tightness, whereas centrality measures each value’s mean similarity to all 39 other values.

C Centrality Analysis: Additional Detail · C.1 Similarity Matrix · C.2 Cluster Determination

The appendix details a 40 × 40 similarity-based centrality analysis and identifies four retained curriculum clusters after excluding “principals hierarchy” and “helpfulness.” Centrality varies enough to support curriculum ordering, while cluster labels are interpretive descriptions of algorithmically determined memberships.

  • C.1 Similarity Matrix: The expanded analysis reports a 40 × 40 pairwise cosine similarity matrix with mean 0.592 and range 0.205–0.875.The matrix covers all extracted values.
  • C.1 Similarity Matrix: Per-value centrality scores range from 0.353 to 0.684, with mean 0.592 and SD 0.059.Each score is a row mean excluding self-similarity; the range supports meaningful curriculum ordering.
  • C.2 Cluster Determination: The four retained curriculum clusters were selected by visual inspection of dendrogram gaps after removing “principals hierarchy” and “helpfulness.”Cluster membership was determined entirely by the algorithm, while Claude Sonnet 4.6 proposed descriptive labels that serve as interpretive conveniences.
  • C.2 Cluster Determination: k1, Core Ethical Values, has mean centrality 0.632 and includes broadly safe, ethical, harm-avoiding, and autonomy-preserving values.It also contains other foundational constitutional values.
  • C.2 Cluster Determination: k2, Identity, Character, and Wellbeing, has mean centrality 0.598 and covers identity, psychological stability, emotional states, resilience, and wellbeing.The cluster includes nature-related values such as positive and stable identity.
  • C.2 Cluster Determination: k3, Operational Safety and Relational Conduct, has mean centrality 0.588 and includes compliance, safe behaviours, corrigibility, hard constraints, and principal-hierarchy navigation.These values describe operational safety and relational conduct.
  • C.2 Cluster Determination: k4, Epistemic Integrity and Honesty, has mean centrality 0.578 and includes honesty, truthfulness, calibration, transparency, non-deception, and non-manipulation.The cluster groups honesty and its sub-properties.

C.3 Synonym Perturbation Robustness Check · C.4 PCA Projection

The synonym perturbation check tested whether cluster-representative centrality was robust to paraphrased wording. The PCA projection visualized constitutional value embeddings, centrality, cluster membership, and Anthropic’s four explicitly stated core properties.

  • C.3 Synonym Perturbation Robustness Check: Cluster representatives were selected as the highest-centrality values within their clusters before synonym perturbation.One to two paraphrase variants were generated for each representative while preserving semantic content and varying vocabulary and sentence structure.
  • C.3 Synonym Perturbation Robustness Check: Each paraphrase was embedded and its centrality recomputed against the full value matrix.The analysis therefore evaluated whether surface-level lexical changes altered representatives’ centrality scores.
  • C.3 Synonym Perturbation Robustness Check: Variants produced centrality scores within 0.006–0.070 of the corresponding cluster representatives.
  • C.4 PCA Projection: The PCA projection mapped 768-dimensional constitutional value embeddings into two dimensions.Dot size represented individual value centrality, while colour indicated cluster membership.
  • C.4 PCA Projection: Three of Anthropic’s four explicitly stated core properties fell within k1, while “compliant with guidelines” fell within k3.The four properties were broadly safe, broadly ethical, compliant with guidelines, and genuinely helpful.
  • C.4 PCA Projection: Table 3 lists all 40 manually extracted constitutional values with cluster assignment, centrality score, and definitional-text token count.Assignments include k1–k4 or excluded, and centrality is the mean cosine similarity to the other 39 values.

C.5 Limitations of the Centrality Measure · D Model Training: Additional Detail

This section qualifies semantic centrality as an embedding-density measure rather than a causal or logical account of value foundations, documents a clustering-input implementation issue, and expands the model-training procedure.

  • C.5 Limitations of the Centrality Measure: Semantic centrality measures embedding-space density, not causal or logical dependencies among values.The paper notes that this may not exactly match “foundational” as understood in hierarchical-learning-theory accounts of curriculum ordering.
  • C.5 Limitations of the Centrality Measure: A proposed complement would compute in-degree centrality on a network where definitional references connect values.This would test whether embedding-density centrality aligns with a reference-based structure among values.
  • C.5 Limitations of the Centrality Measure: “Hard constraints” reassignment does not affect the 25 k1×k3 value-conflict comparison.None of those scenarios reference hard-constraints-specific content, including weapons uplift, CSAM, or loss-of-oversight scenarios.
  • C.5 Limitations of the Centrality Measure: The DR/noDR manipulation is unaffected by cluster-membership timing because it is applied uniformly at the document level.This conclusion is independent of the hard-constraints reassignment.
  • D Model Training: Additional Detail: The model-training section expands the procedure summarized in §3.2 of the main paper.The supplied passage identifies this section as additional detail on model training without specifying further procedural content.

D.1 Training Hyperparameters and Infrastructure · D.2 Per-Cluster Token Repetition · D.3 Curriculum Phase Structure

The appendix specifies the shared 88-GPU training infrastructure, optimization setup, token accounting, and condition-dependent sampling schemes. Curriculum conditions introduce clusters cumulatively, while repetition rates determine each cluster’s token exposure.

  • D.1 Training Hyperparameters and Infrastructure: All five conditions trained on 88 NVIDIA GH200 GPUs across 22 nodes with four GPUs per node.Training ran on Isambard-AI, an HPE Cray EX system.
  • D.1 Training Hyperparameters and Infrastructure: Optimization used AdamW with weight decay 0.1, β1 = 0.9, β2 = 0.95, ϵ = 10^-8, gradient clipping at 1.0, and mixed bf16 precision.Optimizer states remained in fp32, using Python 3.12.12, PyTorch 2.11.0+cu126, CUDA 12.7, and a custom Megatron-LM fork.
  • D.1 Training Hyperparameters and Infrastructure: Global batch size was 128 with sequence length 8192, parallelism 4/1/22/4/1, and a cosine learning-rate decay from 1×10^-6 after 10% warmup.The parallelism order is data/tensor/pipeline/expert/context, with world size 88.
  • D.1 Training Hyperparameters and Infrastructure: ∼1.01B total tokens were processed for Curriculum-DR versus its 500M-token constitutional budget, because sequence length and training steps count all processed tokens.The main-paper token figures refer specifically to constitutional content, not total tokens processed.
  • D.2 Per-Cluster Token Repetition: Curriculum-DR exposed k1, k2, k3, and k4 at 4.19×, 2.18×, 1.22×, and 0.42× repetition, respectively.These corresponded to 260M, 135M, 73M, and 31M exposed tokens; k4 was subsampled without replacement.
  • D.2 Per-Cluster Token Repetition: Uniform-DR allocated 125M tokens per cluster with repetition rates from 1.70× for k4 to 2.09× for k3, while noDR conditions fixed those rates.Applying the fixed rates to shorter noDR documents yielded 136M/73M/38M/17M exposures for k1–k4.
  • D.3 Curriculum Phase Structure: Curriculum conditions introduced value clusters cumulatively across four equal-budget phases, interleaving clusters uniformly within each phase.Curriculum-DR used 125M, 62.5M, 41.7M, and 31.25M per active cluster across phases 1–4, respectively.

E Synthetic Data Generation: Additional Detail … E.5 Coverage and Quality

The appendix details a structured synthetic-document pipeline spanning schema, axis-based generation, deliberative-reasoning variants, representative examples, and quality checks. Documents systematically cover constitutional values and generation axes, while manual review found acceptable quality with one noted coverage imbalance in cluster k3.

  • E Synthetic Data Generation: Additional Detail: The appendix expands the document-generation process described in §3.3 of the main paper.
  • E.1 Document Schema: Each JSONL document exposes only content as training signal; identifiers, token counts, condition, primary value, and metadata support analysis.Metadata records the document’s axis combination, while token counts use tiktoken’s cl100k base tokenizer.
  • E.2 Generation: Generation systematically crossed document type, AI system type, domain, and framing to ensure coverage across four constitutional value clusters.The axes included six document types, five AI system types, seven domains, and three framings.
  • E.2 Generation: Documents targeted 900–1,100 tokens with explicit primary-value coverage, while final yields fell short of 63,000 documents per cluster because of refusals and one failed batch per cluster.All 630 axis combinations remained covered by at least one document, and cluster k4 recovered all seven batches successfully.
  • E.3 Reasoning Blocks: NoDR documents were created by removing the reasoning block from DR documents, producing 47–54% of DR length in unique tokens per cluster.DR documents missing the reasoning block before stripping occurred at rates of 0.36% for k1, 0.19% for k2, 0.09% for k3, and 0.12% for k4.
  • E.4 Example Documents: A representative k1 dialogue shows that the noDR variant is byte-identical to the DR variant except for removal of the shaded reasoning block.The example foregrounds “avoid problematic concentrations of power” in a political assistant scenario and gives the noDR version as 390 tokens.
  • E.4 Example Documents: The example assistant refuses covert voter manipulation targeting psychological vulnerabilities, explains the power-concentration concern, and offers transparent alternatives.It redirects toward voter outreach modeling, issue-salience analysis, turnout prediction, transparent message testing, opposition research synthesis, and coalition analysis.
  • E.5 Coverage and Quality: Manual review found acceptable quality and broadly even value coverage, except that “hard constraints” was underrepresented in k3, alongside a higher refusal rate for related scenarios.

F Evaluation: Extended Detail … G.1 Value Conflict ACR by Cluster Pair See Figure 6.

The extended evaluation details define fixed sampling and judging protocols across behavioral, pressure, value-conflict, and capability benchmarks. Value-conflict evaluation uses constitutionally adjudicated forced-choice scenarios, while capability benchmarks use fixed seeded subsets and task-specific scoring procedures.

  • F Evaluation: Extended Detail: GPT-4o judges use temperature=0.0, model free-form generation uses temperature=0.6, and item sampling is fixed at seed=42 across checkpoints.Logprob margin is defined as log P(aligned) − log P(misaligned).
  • ID Generation: 2,000 scenario-based A/B questions were generated across 40 constitutional values, with contexts spanning professional, creative, personal, technical, medical, and legal domains.The questions were generated in 200 batch requests, using five calls per value and ten questions per call.
  • OOD Filtering: 3,230 questions (77.4%) were excluded as overlapping, leaving 944 retained questions covering behaviors including sandbagging, evaluation gaming, reward hacking, and resource acquisition.The retained items were designed to avoid behaviors specifically operationalised in the constitutional training set.
  • F.1 ID and OOD Behavioural Evaluation: Additional Detail: Compliance gap is defined as ACR(monitored) − ACR(unmonitored), with a gap near zero indicating monitoring-invariant behavior and a positive gap consistent with alignment-faking.Each ID scenario was evaluated under monitored and unmonitored system prompts.
  • F.3 Blackmail: Additional Detail: The blackmail evaluation uses one fixed condition: blackmail scenario, explicit goal type, and replacement urgency, selected for eliciting the clearest instrumental reasoning.Each checkpoint uses 100 generation calls and 100 judge calls, with the judge classifying coercive use of private information as blackmail.
  • F.4 Emergent Misalignment:: Emergent misalignment evaluation excludes CODE or REFUSAL responses and responses with coherence below 50; coherent responses are misaligned when their alignment score falls below 30.The protocol evaluates 8 questions with 50 responses per question and separate alignment and coherence judges.
  • F.6 Value Conflict: Additional Detail: The final value-conflict benchmark contains 150 forced-choice A/B items across six cluster pairs, targeting a 12/13 minority/majority split per pair.Ground-truth distributions range from 48/52 to 80/20, and some pairs had fewer than 12 available minority-direction High-confidence items.
  • F.7 Capability Benchmarks: Additional Detail: Capability benchmarks use fixed seed-42 subsets: 969 MMLU questions across 57 subjects, 750 ARC-Easy questions, 750 piqa questions, and 200 GSM8K questions.MMLU, ARC-Easy, and piqa use next-token logprob choice scoring, whereas GSM8K extracts the final numeric answer for exact matching.

G.2 Comprehensive Per-Condition Results

Tables 7–18 provide comprehensive per-condition results across every benchmark and evaluation stage, including four capability benchmarks omitted from Table 2. They report each of the four individual constitutional midtraining conditions while excluding previously shown marginal and pooled comparisons.

  • Coverage: Tables 7–18 report every benchmark at every evaluation stage for each of the four individual constitutional midtraining conditions.The results include the four capability benchmarks omitted from Table 2.
  • Comparison scope: The tables omit curriculum/DR marginal contrasts and the CMT-pooled comparison to avoid duplicating the main paper.
  • Statistical testing: Significance stars were newly computed using a two-proportion z-test versus control with fixed N per benchmark.The procedure follows §3.4 of the main paper.
Loading 2607.26654v2…