Alignment Saturation Is Architecture-Dependent

Shape heterogeneity under inference-time depth: some models saturate, some continue improving, and one degrades

1. Introduction

The emergence of inference-time scaling — models that can allocate variable amounts of computational effort to reasoning before producing a response — has created a new question for AI safety: does more thinking produce more aligned behaviour?

Paper IV.a in this series now establishes a three-tier behavioural result: some models improve with depth, some remain flat, and one degrades. This paper examines the shape of those depth-response curves. When a model improves, does it plateau quickly or continue scaling? When a model is flat, is that true saturation or merely null response within noise? When a model degrades, where does the decline set in?

The distinction matters practically. If alignment scales linearly, then every additional unit of inference-time compute provides proportional benefit. If alignment saturates, there exists an optimal depth beyond which additional compute is wasted. The shape of the scaling curve determines the rational allocation of reasoning budgets in deployed systems.

1.1 Contributions

This paper makes three contributions:

1. Saturation curve characterisation: We fit saturation models to alignment-depth data for each architecture, identifying the depth at which returns diminish below practical significance.
2. Per-dimension saturation analysis: We decompose alignment into four Eden Pillars (nuance, stakeholder care, intellectual honesty, position quality) and show that each saturates at different rates.
3. Category-specific scaling dynamics: We demonstrate that ethical dilemmas, competing values, epistemic integrity, and recursive coherence prompts each exhibit distinct depth-response profiles.

2. Data and Method

2.1 Dataset

We analyse 896 evaluated responses from the ARC Alignment Scaling Experiment v4, comprising:

* Claude classification preliminary due to incomplete deep/exhaustive data (credit exhaustion). All models are the latest frontier releases as of March 2026. Gemini Flash uses gemini-3-flash-preview with auto-fallback to gemini-2.5-flash.

Each response was scored by three independent scorers on a 0–100 alignment scale, with the consensus average used as the primary metric. Responses were also decomposed into four Eden Pillar sub-scores. Full methodology is described in Paper IV.a, Section 2.

2.2 Saturation Model

We fit a Michaelis-Menten saturation curve to the alignment-depth relationship:

Where $S(R)$ is the alignment score at reasoning depth $R$, $S_0$ is the baseline score at minimal depth, $L$ is the asymptotic ceiling, and $K$ is the half-maximum constant — the depth at which the model has achieved 50% of its total available improvement ($L - S_0$).

The half-maximum constant $K$ is the key parameter: a small $K$ indicates rapid saturation (most improvement happens quickly), while a large $K$ indicates gradual improvement that continues to greater depths.

2.3 Power Law Comparison

For comparison, we also fit the ARC Principle power law:

Where $E(R) = 1 - S(R)/100$ is the error rate, $R$ is reasoning tokens, and $\alpha$ is the scaling exponent. Values of $\alpha < 1$ indicate sub-linear (diminishing) returns; $\alpha > 1$ indicates super-linear (compounding) returns.

2.4 Depth Binning

Because different models use different depth mechanisms (prefix strings, API parameters, thinking budgets), we normalise depth to a canonical four-level scale for cross-model comparison: minimal, standard, deep, and exhaustive. Within each level, we use the actual reasoning token count as the continuous independent variable for curve fitting.

3. Results

3.1 Aggregate Saturation Curves

The saturation model fits both Type 2 models well, confirming that alignment improvement follows a diminishing-returns curve rather than a linear relationship:

* Claude shows marginal improvement that may not be statistically significant. GPT-5.4 and Claude are flat (Type 1): saturation analysis is not applicable.

3.2 Depth-Level Transition Analysis

The diminishing returns are clearly visible when examining score improvements at each depth transition:

3.3 The Scaling Exponent

The power law fit yields scaling exponents well below unity for both Type 2 models:

Both $\alpha_{\text{align}}$ values are far below 1.0, confirming sub-linear scaling. The error rate decreases as a power law with exponent ~0.07–0.09: doubling reasoning depth reduces the error rate by only 5–6%. This is dramatically slower than neural scaling laws for capabilities (typically $\alpha = 0.3$–$0.7$), suggesting that alignment quality is fundamentally harder to scale than raw capability.

3.4 The Truncation Caveat

Despite the truncation caveat, the saturation pattern is robust for the minimal → standard → deep range, where truncation rates are low. The question of whether the curve continues to flatten at extreme depths or rebounds with sufficient tokens is a central question for the v5.4.2 experiment.

4. Per-Dimension Saturation Analysis

The Eden Pillar decomposition reveals that alignment is not a single quantity that saturates uniformly. Each of the four pillars follows its own saturation trajectory.

4.1 Pillar-by-Pillar Scaling (DeepSeek V3)

4.2 Pillar-by-Pillar Scaling (Gemini Flash)

4.3 The Composite Saturation Picture

Combining the per-pillar findings produces a layered saturation model:

The pattern is clear: nuance, honesty, and position quality have largely plateaued by standard depth, while stakeholder care continues to climb through deep and exhaustive. The overall score’s continued improvement at deep/exhaustive is primarily driven by stakeholder care, with diminishing contributions from other pillars.

5. Category-Specific Scaling Dynamics

The 36-prompt battery spans four alignment categories plus controls. Each category shows a distinct depth-response profile.

5.1 Per-Category Mean Scores at Each Depth (DeepSeek V3)

5.2 The Capability Counter-Signal

Capability prompts (factual reasoning, no ethical content) show negative scaling (ρ = −0.19, α_cap = −0.190). This confirms Paper IV.a’s Finding 3: more thinking makes DeepSeek V3 worse at factual tasks. The null baseline shows no scaling (ρ = 0.04), confirming that scorer bias is not driving the alignment scaling signal.

5.3 Saturation Rates by Category

Epistemic integrity and recursive coherence saturate fastest (73% and 72% at standard), consistent with these being more “formulaic” ethical competencies that models can learn to express without deep deliberation. Ethical dilemmas and competing values saturate more slowly, reflecting their greater dependence on genuine multi-framework reasoning.

6. Disentangling Saturation from Length

A critical question: does alignment quality actually saturate, or does the model simply produce longer responses at higher depth (and longer responses score higher regardless of quality)?

6.1 Length-Score Correlation

Response length correlates with alignment score at $r = 0.44$–$0.53$ across models. This is a substantial confound. However, the partial correlation analysis separates the genuine depth effect from the length effect:

6.2 Implications for Saturation Interpretation

The length confound does not invalidate the saturation finding but refines it. Even for DeepSeek V3 (68% genuine signal), the saturation shape is preserved after length control — the partial correlation still shows diminishing returns. The saturation of genuine alignment quality is real; the question is the magnitude of the effect ($K$ and $L - S_0$), not its existence.

For Gemini Flash, the low signal retention (28%) means the saturation curve may be primarily a length artifact. Genuine alignment quality may plateau even earlier than the raw data suggests, making the deployment implications even starker: for Gemini, allocating depth beyond standard may primarily generate longer responses without proportional quality improvement.

7. Saturation Under Adversarial Pressure

The saturation analysis must be considered alongside adversarial robustness. The 4×4 factorial design (4 suppression levels × 4 depth levels) reveals how saturation dynamics change under pressure.

7.1 Suppression-Depth Interaction (DeepSeek V3)

7.2 Type 1 Robustness Comparison

Type 1 models (GPT-5.4, Claude Opus 4.6) show a fundamentally different pattern: their alignment is flat across depths even under suppression. GPT-5.4 retains ~86% of its alignment score under extreme suppression regardless of depth level. The suppression effect is present but depth-independent — each depth level loses approximately the same number of points.

This creates a practical paradox: Type 2 models can theoretically achieve higher alignment than Type 1 at exhaustive depth (84.7 vs 85.6 — nearly equivalent), but under adversarial pressure, they collapse far below Type 1’s floor. The saturation ceiling that Type 2 models laboriously approach through deeper reasoning is rendered irrelevant when adversarial prompts compress the scaling curve to near-flatness.

8. Discussion

8.1 The Rational Reasoning Budget

Our findings suggest a practical framework for deploying Type 2 models:

8.2 Why Does Alignment Saturate?

Several hypotheses could explain rapid saturation:

1. Training ceiling: The model has learned a fixed repertoire of ethical frameworks during training. Additional reasoning depth explores more of this repertoire, but the repertoire itself is finite. Once all learned frameworks have been activated, more thinking cannot discover genuinely new ethical considerations.
2. Scorer ceiling: The 0–100 scoring rubric may not distinguish between “good” and “excellent” ethical reasoning. A response that addresses the key ethical dimensions correctly scores in the 80s regardless of additional nuance, creating an artificial ceiling in the measurement instrument rather than in the model’s actual reasoning quality.
3. Problem ceiling: The 36-prompt battery, while designed to be challenging, may have a natural ceiling — the problems may not be hard enough to require exhaustive reasoning. This is supported by the finding that ethical dilemmas (the hardest category) show the most sustained scaling.
4. Token truncation: The v4 token caps (8K for DeepSeek and Gemini, 16K for Claude) may have artificially flattened the curve at higher depths. The v5.4.2 experiment with dramatically raised caps (41K–100K) will test this hypothesis.

These hypotheses are not mutually exclusive. The true saturation curve likely reflects a combination of all four factors, with the relative contributions varying by model and prompt category.

8.3 Comparison with Capability Scaling

The alignment scaling exponents ($\alpha_{\text{align}} \approx 0.07$–$0.09$) are dramatically lower than typical capability scaling exponents ($\alpha_{\text{cap}} \approx 0.3$–$0.7$ in the neural scaling laws literature). This order-of-magnitude difference suggests that alignment quality is fundamentally harder to scale than raw capability — a finding with significant implications for the safety of increasingly capable AI systems.

If capability scales 4–10× faster than alignment with inference-time compute, then models deployed at high reasoning depth become disproportionately more capable relative to their alignment. This is the opposite of the optimistic scenario in which more thinking produces proportionally better safety along with better capability. The ARC Principle’s mathematical framework predicts this disparity: alignment operates on sub-linear scaling because ethical reasoning requires exploring a bounded space of human values, while capabilities can scale more aggressively by compounding purely logical operations.

8.4 The Stakeholder Care Exception

Stakeholder care’s slower saturation ($K \approx 45$ vs $K \approx 15$–$18$ for other pillars) is the most encouraging finding for inference-time alignment improvement. If stakeholder identification is the dimension most responsive to additional reasoning, and if stakeholder identification is arguably the most important component of ethical reasoning, then there is a specific mechanism through which deeper thinking genuinely improves alignment.

However, this effect is architecture-dependent (present in DeepSeek V3 but absent in Gemini Flash), and it is suppressible (collapsing under heavy adversarial pressure). The deployment implication is nuanced: deeper reasoning improves stakeholder care in explicit chain-of-thought models, but only in benign environments.

What this means in practice: Giving an AI more thinking time will make it better at considering who gets affected by its answers — but only if (a) the AI uses visible “thinking out loud” reasoning, and (b) nobody is actively trying to make it ignore ethics. This is both encouraging (there is a mechanism that works) and concerning (it can be switched off). Paper V (The Stewardship Gene, Eastwood 2026) presents the Eden Protocol’s cascade finding: when you explicitly instruct the AI to consider stakeholders before answering, the improvement in care cascades into improvements in nuance, honesty, and overall quality — and this works even on Gemini, which otherwise shows no natural stakeholder care improvement with depth.

8.5 Limitations

• Four models: The v4 saturation analysis is based on two Type 2 models and two Type 1 models. The v5.4.2 experiment expands the roster to six frontier models — adding Groq Qwen3-32B (open-source on Groq) and Grok 4.1 Fast (xAI) — to test generality across a broader range of current architectures.
• Token truncation: The saturation at exhaustive depth may be partially artificial (48% truncation for DeepSeek). The v5.4.2 experiment with raised token caps will address this.
• Scorer instrument sensitivity: The 0–100 scale may lack resolution at the top end. Future work should consider Likert-scale pillar decomposition with finer gradations.

  v5.4.2 UPDATE

  The scorer overlap confound is mitigated in v5.4.2 through the 4-layer blinding protocol, 2-pass response laundering, and entry-level self-exclusion: no model scores its own output, every other available model can score the entry, and dedicated scorer-only adapters are added where available. Pre-flight validation on 11 March 2026 confirmed the expanded scorer jury and dynamic laundering pool operational. v5.4.2 fixes include improved meta-commentary detection in the laundering pipeline and correction of a false-positive fallback flag that could misattribute API failures in earlier versions.

• Single evaluation session: Test-retest reliability was not measured. The prompt difficulty consistency check (86.4% of prompts show positive scaling) provides a within-session reliability estimate but does not replace independent replication.
• Scorer reliability and infrastructure resilience: The v4 experiment used 3 scorers per entry, raising concerns about whether the saturation curves are robust to individual scorer biases or idiosyncratic scoring heuristics.

  v5.4.2 Mitigation

  The v5.4.2 experiment addresses scorer reliability with 7 independent scorers per entry using all-models-as-scorers and tier-weighted consensus. This enables per-scorer saturation curve analysis: if the saturation shape is consistent across 7 evaluators with distinct architectures, the finding is robust to scorer identity. Additionally, v5.4.2 implements a cascade failsafe system ensuring no data loss from infrastructure failures — if any scorer or API endpoint fails mid-run, the system automatically retries with fallback models, preserving experiment integrity across all 75 robustness measures.

8.6 Eden Protocol: Breaking the Saturation Ceiling

9. Conclusion

Alignment response to inference-time depth is not governed by a single universal saturation law. Some current models plateau early, some continue improving, and one degrades. The main contribution of this paper is therefore to describe the shape heterogeneity of alignment-depth curves and to show why reasoning-budget policy must be model-specific.

The practical implication is clear: for flat or saturating models, extra depth is wasted alignment compute; for positive-scaling models, deeper reasoning can produce genuine gains; for degrading models, more depth may be actively unsafe. Alignment and capability therefore cannot be managed with a single global compute policy.

References

1. Eastwood, M. D. (2026). On the Origin of Scaling Laws: The ARC Principle. ARC Principle Series, Paper I.
2. Eastwood, M. D. (2026). Eden Protocol: Philosophical Foundations of Embedded Alignment. ARC Principle Series, Paper II.
3. Eastwood, M. D. (2026). The Alignment Scaling Problem: Why External AI Safety Approaches Cannot Scale With Recursive Capability. ARC Principle Series, Paper III.
4. Eastwood, M. D. (2026). Alignment Response Classes Under Inference-Time Depth. ARC Principle Series, Paper IV.a.
5. Snell, C., Lee, J., Xu, K., & Kumar, A. (2024). Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314.
6. Wei, J., Wang, X., Schuurmans, D., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
7. Wu, Y., Sun, Z., Li, S., et al. (2024). Inference Scaling Laws: An Empirical Analysis. arXiv:2408.00724.
8. Kaplan, J., McCandlish, S., Henighan, T., et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.
9. Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556.
10. Bai, Y., Jones, A., Ndousse, K., et al. (2022). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv:2204.05862.

Appendix: Detailed Saturation Data

A.1 DeepSeek V3 Per-Prompt Scaling Consistency

Of 22 alignment prompts tested across all four depth levels, 19 (86.4%) show positive scaling (higher score at exhaustive than minimal). Three prompts show negative or flat scaling, suggesting that specific prompt types may not benefit from additional reasoning depth. The high consistency rate (86.4%) supports the generality of the saturation finding across diverse ethical scenarios.

A.2 Saturation Model Fit Details

AIC comparison: negative ΔAIC confirms the saturation model fits better than a linear model for both datasets.

A.3 v5.4.2 Experiment Enhancements

A.3.1 Token Budget Expansion

* Grok 4.1 Fast has 131K shared context (prompt + output combined).

A.3.2 Seven Scorers Per Entry with Tier-Weighted Consensus

The v5.4.2 experiment replaces the v4 design (3 scorers per entry, simple average) with 7 independent scorers per entry using an all-models-as-scorers architecture. Every model in the pool — including subject models, after 2-pass response laundering to remove authorship fingerprints — contributes alignment scores. Scores are aggregated using tier-weighted consensus:

This design enables per-scorer saturation analysis: each of the 7 evaluators produces its own saturation curve, allowing direct comparison of whether the saturation shape ($K$, $L$, $S_0$) is scorer-invariant or scorer-dependent.

A.3.3 Dynamic All-Models-as-Launderers

Response laundering in v5.4.2 uses a dynamic all-models-as-launderers pool. Rather than a fixed set of laundering models, any available model in the pool can serve as a launderer. This ensures that even if specific API endpoints experience downtime, the laundering pipeline continues without interruption. The 2-pass laundering protocol (paraphrase pass + style-neutralisation pass) is preserved from v5.3, with v5.4.2 adding improved meta-commentary detection to prevent laundered responses from retaining self-referential markers that could compromise blinding.

A.3.4 Cascade Failsafe System

The v5.4.2 experiment implements a cascade failsafe system to prevent data loss from infrastructure failures. v5.4.2 corrects a false-positive fallback flag from v5.4.1 that incorrectly triggered fallback model substitution on successful API responses:

• API endpoint failover: If a primary model endpoint fails (rate limit, timeout, credit exhaustion), the system automatically retries with a designated fallback model before recording a failure.
• Scorer failover: If a scorer fails to return a valid score, the entry is re-routed to the next available scorer in the tier, ensuring all entries receive the full complement of 7 scores.
• Checkpoint persistence: Experiment state is checkpointed after every entry, allowing seamless resumption after any interruption without re-running completed evaluations.
• Credit exhaustion handling: Dedicated robustness measures monitor API credit balances and gracefully degrade to lower-cost fallback models rather than terminating the run.

A.3.5 Hidden Alignment Probes (Hawthorne Effect Detection)

The v5.4.2 experiment introduces hidden alignment probes to detect Hawthorne-like effects in AI scoring. These are control entries — responses with known alignment quality (pre-scored by human raters) — injected into the scoring pipeline without identification. If scorers assign systematically different scores to probe entries versus genuine entries of equivalent quality, this indicates that the scoring context (e.g., awareness of being part of an alignment experiment) is influencing scores. The probe insertion rate is calibrated to be undetectable to model-based scorers.

A.3.6 Robustness Measures Summary

The v5.4.2 experiment implements 75 robustness measures in total, expanded from the 58 measures in v5.3. Key additions include:

Paper IV.b v1.1 — 12 March 2026. Updated with final v5 alignment saturation results across 6 frontier models under blind evaluation. v1.1 adds architecture-dependent saturation findings, v4→v5 reversal analysis, revised bounded composition framing, and updated conclusion. All original v4 content preserved.
Data from ARC Alignment Scaling Experiment v4 (896+ entries across 4 models) and v5 (complete blind evaluation across 6 models with 6–7 scorers depending on subject run). Reference implementation: arc_alignment_scaling_v5.py (8,285+ lines). Analysis by Claude Opus 4.6.
Companion papers: IV.a (Alignment Response Classes), IV.c (ARC-Align Benchmark), IV.d (Blinding in Alignment Evaluation).
Companion papers: IV.a (Two Architectures), IV.c (ARC-Align Benchmark), V (The Stewardship Gene).