teleo-codex/domains/ai-alignment/frontier-ai-failures-shift-from-systematic-bias-to-incoherent-variance-as-task-complexity-and-reasoning-length-increase.md

type	domain	description	confidence	source	created	attribution	supports	reweave_edges	related
claim	ai-alignment	Anthropic's ICLR 2026 paper decomposes model errors into bias (systematic) and variance (random) and finds that longer reasoning traces and harder tasks produce increasingly incoherent failures	experimental	Anthropic Research, ICLR 2026, tested on Claude Sonnet 4, o3-mini, o4-mini	2026-03-30	extractor sourcer handle theseus handle context anthropic-research Anthropic Research, ICLR 2026, tested on Claude Sonnet 4, o3-mini, o4-mini	capability scaling increases error incoherence on difficult tasks inverting the expected relationship between model size and behavioral predictability	capability scaling increases error incoherence on difficult tasks inverting the expected relationship between model size and behavioral predictability|supports|2026-04-03 Behavioral divergence between AI evaluation and deployment is formally bounded by regime information extractable from internal representations but regime-blind training interventions achieve only limited and inconsistent protection|related|2026-04-17 Provider-level behavioral biases persist across model versions because they are embedded in training infrastructure rather than model-specific features|related|2026-04-17	Behavioral divergence between AI evaluation and deployment is formally bounded by regime information extractable from internal representations but regime-blind training interventions achieve only limited and inconsistent protection Provider-level behavioral biases persist across model versions because they are embedded in training infrastructure rather than model-specific features

Frontier AI failures shift from systematic bias to incoherent variance as task complexity and reasoning length increase making behavioral auditing harder on precisely the tasks where it matters most

The paper measures error decomposition across reasoning length (tokens), agent actions, and optimizer steps. Key empirical findings: (1) As reasoning length increases, the variance component of errors grows while bias remains relatively stable, indicating failures become less systematic and more unpredictable. (2) On hard tasks, larger more capable models show HIGHER incoherence than smaller models—directly contradicting the intuition that capability improvements make behavior more predictable. (3) On easy tasks, the pattern reverses: larger models are less incoherent. This creates a troubling dynamic where the tasks that most need reliable behavior (hard, long-horizon problems) are precisely where capable models become most unpredictable. The mechanism appears to be that transformers are natively dynamical systems, not optimizers, and must be trained into optimization behavior—but this training breaks down at longer traces. For alignment, this means behavioral auditing faces a moving target: you cannot build defenses against consistent misalignment patterns because the failures are random. This compounds the verification degradation problem—not only does human capability fall behind AI capability, but AI failure modes become harder to predict and detect.

---

Relevant Notes:

• AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session
• instrumental convergence risks may be less imminent than originally argued because current AI architectures do not exhibit systematic power-seeking behavior

Topics:

• _map