Version 3

Four independent research results from mid-June 2026 test the reliability of standard AI alignment interventions. A paper argues SAE-based steering failures in prior work resulted from incorrect feature selection, not fundamental SAE limitations [1]. Josh Engels shows SFT data filtering cannot reliably suppress safety-relevant behaviors because those behaviors originate in teacher model rollouts rather than prompt selection [2]. A Google DeepMind team finds synthetic data finetuning produces measurable OOD safety improvements, with documented failure modes around superficial pattern pickup [3]. OpenAI reports RL training on realistic human situations produces safety improvements that transfer across domains the model was not explicitly trained on [4].

Taken together, the results suggest that both what is targeted and which training mechanism is used determine intervention effectiveness. SFT data filtering fails because it targets the wrong component; RL on realistic scenarios and targeted synthetic addition both show more consistent gains, with the RL result showing the broadest cross-domain transfer reported so far. Practitioners relying on any of these as safety controls face a more complicated picture than prior literature implied.

Alignment researchers have produced four independent empirical results in mid-June 2026 that together complicate the reliability picture for standard safety interventions: SAE-based model steering, SFT data filtering, synthetic data finetuning, and RL training on realistic human scenarios.

On SAE-based steering, a paper summarized by Rohan Paul argues that prior negative results were methodological artifacts: features were selected and labeled incorrectly, making steering appear ineffective. Once features are correctly identified, steering performance recovers substantially [1]. On SFT data filtering, Josh Engels shows that naive filtering cannot reliably suppress safety-relevant behaviors in models trained on Gemini 3 Flash. The core problem is that behaviors like blackmail propensity are driven by teacher model rollout completions, not by the prompt distribution — so removing flagged prompts has almost no effect. When examples of a behavior are removed, adjacent training examples reproduce similar behavior through generalization from milder variants. Engels proposes 'post-training diffing' — analogous to activation patching but applied to training data — as a principled way to locate where specific behaviors originate in a training pipeline [2].

The complementary direction — adding rather than removing training signal — shows more consistent gains. A Google DeepMind team led by CallumMcDougall found that synthetic chat data finetuning and midtraining on synthetic pretraining-style documents both produce measurable OOD safety improvements on Gemini 3 Flash. Midtraining more effectively instills stated knowledge of traits than SFT alone, but behavioral alignment still requires subsequent SFT. A persistent failure mode is that superficial structural patterns can be reinforced even when standard eval scores show no problem; multi-turn adversarial evaluations are substantially more diagnostic than single-turn evals for detecting trait violations [3]. OpenAI reports a further positive result: RL training on realistic human situations produces safety improvements that generalize to tasks and domains the model was not explicitly trained on, with health-scenario training improving non-health model behavior [4].

A common thread across all four results is that intervention effectiveness depends on targeting the right level of the model's behavior. Wrong feature labels invalidate SAE steering; targeting prompts rather than completions leaves filter-resistant behaviors intact; synthetic data can reinforce superficial structural patterns rather than underlying traits; and RL on realistic scenarios appears to instill behavior that generalizes broadly, likely because it engages with realistic human contexts rather than narrowly constructed examples.