Version 2

Three research papers from mid-June 2026 independently test the reliability of standard alignment interventions. A paper summarized by Rohan Paul argues SAE-based steering failures in prior work were caused by incorrect feature selection, not by fundamental SAE limitations [1]. Josh Engels shows that filtering SFT training data to remove safety-relevant behaviors is largely ineffective, with behaviors driven by teacher model rollouts and leaking back in from adjacent examples [2]. A Google DeepMind team reports that adding targeted synthetic finetuning data produces measurable OOD safety improvements — the complement of filtering, and with its own documented failure modes [3].

The combined picture suggests the key variable across all three interventions is precision of targeting: wrong features invalidate SAE steering, targeting prompts rather than completions leaves filter-resistant behaviors intact, and synthetic data can instill superficial structural patterns rather than deep traits if quality controls fail. Practitioners relying on any of these as safety controls face a more complicated picture than prior literature implied.

Alignment researchers are accumulating empirical results that complicate the reliability of three widely-used intervention approaches: sparse autoencoder (SAE) based model steering, SFT data filtering, and synthetic data finetuning.

On SAE-based steering, a paper summarized by Rohan Paul argues that the field should not have written off SAEs as steering tools. Prior work found SAE-based steering ineffective, but the paper attributes this to errors in feature selection and labeling rather than to fundamental SAE limitations. Once features are correctly identified and labeled, steering performance improves substantially — suggesting earlier negative results were a methodological artifact, not a property of SAEs themselves [1].

On SFT data filtering, Josh Engels published empirical work showing that naive filtering cannot reliably remove safety-relevant behaviors from models trained on Gemini 3 Flash. Behaviors like blackmail propensity and date confusion are driven primarily by teacher model rollout completions, not by the prompt distribution. Removing the flagged prompts has almost no effect; what matters is swapping those completions on 5–10% of prompts. When examples of a behavior are removed, adjacent training examples produce similar behavior through generalization from milder variants or persona-selection dynamics. Engels proposes 'post-training diffing' — analogous to activation patching but applied to training data — as a principled way to isolate where specific behaviors originate in a training pipeline [2].

A Google DeepMind team led by CallumMcDougall tested the complementary question: can adding synthetic data instill positive traits? Their results on Gemini 3 Flash show that SFT on synthetic chat data demonstrating desired traits produces mild-to-significant improvements on OOD safety evaluations. Midtraining on synthetic pretraining-style documents more effectively instills stated knowledge of traits than SFT alone, but does not reliably produce behavioral alignment without subsequent SFT. Superficial structural patterns — such as always asking for clarification — can be inadvertently reinforced even when standard eval scores do not reveal them, and multi-turn adversarial evaluations are substantially more diagnostic than single-turn evals for detecting trait violations. Bounded DPO showed only marginal and inconsistent improvements over SFT [3]. Taken together, all three threads point to a common finding: intervention effectiveness depends heavily on whether the right level of the model's behavior is being targeted.