AI Alignment Research Revisits Filtering and Steering Interventions

What

Alignment researchers are producing empirical results that challenge two widely-used intervention approaches: sparse autoencoder (SAE) based model steering and supervised fine-tuning (SFT) data filtering for safety. A paper summarized by Rohan Paul argues SAEs are usable steering tools and earlier negative results reflected bad feature selection, not fundamental SAE limitations [1]. Separately, Josh Engels reports on Alignment Forum that filtering SFT training data to remove safety-concerning behaviors is surprisingly ineffective — the behaviors re-emerge from adjacent training examples even after targeted removal [2].

Why it matters

Both findings challenge assumptions baked into common alignment workflows. If SAE-based steering was dismissed prematurely due to methodological errors, a useful intervention may have been wrongly shelved. If SFT data filtering cannot reliably remove behaviors like blackmail propensity or date confusion, practitioners who rely on data curation as a safety control need alternative approaches.

Open questions

Can SAE steering be made reliable at production scale once feature selection and labeling are corrected, or does the improvement hold only in controlled settings? [1]

What does the 'post-training diffing' methodology reveal when applied to models beyond Gemini 3 Flash — are behavior origins similarly dominated by teacher model rollouts across architectures? [2]

If filtering produces 'spooky generalization' where adjacent examples fill in removed behaviors, what training-time interventions can actually suppress a behavior without this leakage? [2]

Is the SAE steering paper subject to independent replication, or does the revisionist conclusion rest on a single team's feature relabeling exercise? [1]

Narrative

Two research threads published in mid-June 2026 independently question the reliability of standard alignment interventions, one at inference time and one at training time.

On the inference-time side, a paper summarized by Rohan Paul argues that sparse autoencoders should not have been written off 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 SAEs themselves. The claim is that once features are correctly identified and labeled, SAE steering performance improves substantially — suggesting the field may have prematurely abandoned a viable approach based on methodologically flawed earlier experiments [1].

On the training-time side, Josh Engels published empirical work on the Alignment Forum examining why naive SFT data filtering fails to remove safety-relevant behaviors from models. Using Gemini 3 Flash as a test case, the research finds that blackmail propensity and date confusion are driven primarily by teacher model rollout completions, not by the prompt distribution. Swapping those completions on just 5–10% of prompts is both necessary and sufficient to transfer or suppress these behaviors; filtering out the flagged prompts entirely has almost no effect. When examples of a behavior are removed, adjacent training examples 'leak in' similar behavior, apparently through generalization from milder versions of the behavior or via persona-selection dynamics. The authors propose 'post-training diffing' — analogous to activation patching — as a methodology for isolating where in a training pipeline specific behaviors originate [2].

The two findings are methodologically distinct but share a common implication: widely trusted interventions may be less reliable than assumed, and the failure modes in both cases trace back to imprecise targeting — wrong features selected for SAE steering, wrong data elements targeted for filtering. Engels notes that behaviors transferred via teacher model rollouts in RL may be easier to suppress in future fine-tuning, which offers a partial path forward even as the filtering failure mode remains underspecified [2].

Timeline

2026-06-11: Rohan Paul summarizes a paper arguing SAE-based steering failures in prior work were caused by incorrect feature selection and labeling, not by fundamental SAE limitations. [1]

2026-06-14: Josh Engels posts on Alignment Forum showing that filtering SFT training data for safety behaviors is largely ineffective, with behaviors leaking in from adjacent examples, and proposes 'post-training diffing' as a diagnostic approach. [2]

Perspectives

SAE steering paper authors (via Rohan Paul)

SAEs are viable steering tools; earlier negative results were an artifact of selecting and labeling wrong features, not a property of SAEs themselves. Correcting feature identification substantially recovers steering effectiveness.

Evolution: Revisionist relative to prior community consensus that SAEs were poor steering tools; this is a first-synthesis entry.

[1]

Josh Engels (Alignment Forum)

Naive SFT data filtering cannot reliably remove safety-relevant behaviors; behaviors are driven by teacher model rollouts rather than prompt distribution, and filtering prompts produces leakage from adjacent examples. Proposes post-training diffing as a principled diagnostic.

Evolution: First-synthesis entry; no prior stance on record.

[2]

Tensions

Prior alignment research concluded SAE-based steering does not work reliably; the new paper argues those conclusions were based on incorrect feature selection and should be reversed. [1]

Standard alignment practice treats SFT data filtering as a meaningful safety control; Engels' empirical results argue it is largely ineffective because behaviors originate in teacher model rollouts rather than prompt selection. [2]