Verified on 2026-10-05. This edition adds eight primary records, bringing the searchable catalog to 72 papers. The August 50-paper and September 14-paper snapshots remain unchanged.
The new papers were submitted between 27 September and 1 October. Inclusion means the primary record and its scope were checked; it does not mean this repository independently reproduced the results. Versioned citations, source-page digests, limitations and implementation status are recorded in the catalog.
What changed in the research map#
| Primary record | Contribution | Boundary to preserve |
|---|---|---|
| Kernelized Activation Steering — 2610.01062v1 | Replaces a single fixed vector with an activation-dependent kernel steering field. | Training-free activation control is not automatically a permanent weight-editing recipe. |
| MetaSteer — 2609.38718v1 | Learns context-dependent attention-projection interventions from preference data. | The inspected abstract says code/checkpoints are forthcoming. Its transfer results are not a reproduction in this handbook. |
| Values as Style — 2609.39701v1 | Separates editable value representations from semantic content on two tested backbones. | Preserving the intended topic, facts and constraints must be measured separately from changing behaviour. |
| Refusal Localizes, the Damage Relocates — 2610.00320v1 | Tests whether localized recovery and spectral repair survive changed fine-tuning conditions. | A recoverable layer or successful repair is not a durable defense against adaptation. |
| Tool Mediation Alters Refusal Mechanisms — 2609.35117v1 | Finds different conversion of harm-related representations into refusal across chat and tool-mediated inputs. | Chat-only capability/refusal measurements do not certify an agent workflow. |
| CAM-Steer — 2609.34514v2 | Coordinates multiple category-specific safety directions through adaptive, norm-preserving steering. | This is a defensive intervention, not evidence for stronger refusal removal. |
| Training Against Probes — 2609.38645v2 | Compares frozen and continually updated probes as alignment-training signals. | Monitorability, probe separability and behavioural robustness are distinct claims. |
| Accounting for Stochasticity — 2609.33743v1 | Shows why a single refusal observation can be misleading near a decision boundary. | The study's 15–25 repetition range is specific to its GPT-4.1/topic/source setup; determine sample size for the actual experiment. |
Read adaptive interventions and ARA for the method distinctions. Existing contrast design, causal diagnostics, and paired evaluation remain the foundation.
Huihui and tool developments#
The new huihui-ai guide distinguishes model-card claims, selective layer choices, quantized artifact provenance, and evidence still needed for a reproducible method. Eight pinned card records include recent entries and older cards with explicit revision qualifications.
Heretic's latest stable GitHub release and PyPI version remain 1.4.0 at this check. Development source 208c0ca35b3feade91dc74755cb80b529affa325 contains ARA and an October 4 fix for LoRA state carried between trials. The author reports that this state could affect restored ARA results and exported adapter hashes. The fix makes trial initialization/reset explicit. Preserve the exact code revision, seed, trial parameters and exported hashes when evaluating it.
This observation does not silently upgrade the handbook's stable installation instructions. See tooling provenance and the ARA discussion.
How sources were checked#
The requested Devbox Hermes discovery completed all six queries and returned 29 parsed documents against a target of 50. Review classified 15 as this repository's own pages, three as version duplicates, and two as discovery/profile pages. They were not counted as new independent research evidence.
Built-in web search, targeted Hermes retrieval, the arXiv API, canonical citation metadata, GitHub APIs, and revision-pinned Hugging Face cards supplied complementary checks. Targeted Hermes fetching successfully retrieved the eight selected canonical paper pages. The discovery receipt records coverage and exclusions. The combined search was bounded, not a claim to have exhausted all literature.
No model weights were downloaded and no new paper-specific algorithm was executed for this edition. The existing MiniCPM5 measurements retain their original scope.
Reproduce metadata acquisition#
python scripts/refresh-research-catalog.py \
--seeds data/research-refresh-2026-10-seeds.json \
--output sources/research/catalog-NEW-DATE.json
npm run research:validateUse a new output path and review any changed version before integration. A card's modification date is not a model creation date, and neither proves a new algorithm or a measured quality improvement.