Heretic 2.0 rewrites model-modification tool with arbitrary-rank ablation and plugins

Heretic 2.0 replaces the method at the center of the open-source model-modification tool. Its default modifier is now Arbitrary-Rank Ablation (ARA), and the release breaks compatibility with 1.x configuration files. Heretic is built to remove refusals from language models without training them, and the project describes the approach as training-free. The change matters because it alters how the edit is computed, not only how it is configured.

Heretic 1.x relied on directional abliteration, which the README calls the classic method for refusal suppression. That method removes a single direction from the model's internal representations, which is what the release means by rank-1. ARA instead allows edits of arbitrary rank, so the change to each weight matrix can span more than one direction. A Reddit discussion of the release describes the goal as making activations on harmful prompts resemble those on harmless prompts while minimally perturbing harmless prompts. The summary of the release describes the per-matrix fitting as L-BFGS optimization over residual-space representations. The README adds that Heretic combines the modifier with a TPE-based hyperparameter optimizer powered by Optuna, which searches for parameters that minimize or maximize the configured scores.

Those scores come from scorer plugins, which are the second major change. By default Heretic scores models on refusal rate, using keyword matching, and on KL divergence, which measures how far the modified model drifts from the original. Several scorers can run at once and are co-optimized, so the optimizer is trading one objective against another. The Reddit discussion lists further examples, including benchmark scoring and a ModernBERT-based refusal and emotion classifier. A plugin can be loaded either by a fully qualified import path or by a file path followed by a class name, so users can add scorers without modifying the package. Modifier plugins, which implement steering methods, are the counterpart. The README notes, however, that only one modifier can be active per run for now.

The compatibility break is concrete. The migration guide lists research settings that are no longer recognized, including residual plot options. It removes the kl_divergence_scale and kl_divergence_target settings because KL balancing was dropped. Abliteration settings move under a [modifier.Abliteration] section, refusal_markers is renamed keyword_markers, and the good evaluation prompts become the KLDivergence prompts. Because 2.x defaults to ARA, a user who wants the old behavior has to set the modifiers option explicitly. The guide says only settings present in an old file need changing.

Processing Qwen3-4B-Instruct-2507 on an RTX 3090 with the default configuration takes about 30 to 40 minutes. The project says the tool benchmarks the system at startup to choose a batch size. It also says that thousands of models have been published with Heretic, and that independent benchmarks on MMLU and GSM8K found them comparing favorably with models from competing abliteration tools. The README gives no head-to-head result for ARA against the old abliteration default. Reproducibility manifests are a separate feature, and the README says each reproduction manifest is only around 9 kilobytes.

Community reaction in the Reddit thread was mostly enthusiastic, with users especially interested in using the plugin system and ARA to reduce what they call slop. A separate thread raised attribution concerns about a different tool, reaper-abliteration, which a forensic write-up links to the developer HauhauCS. The write-up argues that the closed-source package is a fork of Heretic that was relicensed to PolyForm Noncommercial and possibly refactored with an LLM. It cites a statement from Heretic's author, Philipp Emanuel Weidmann, who says the package was plagiarized from Heretic. The write-up also alleges violations of the AGPL-3.0 license that Heretic uses.