Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities
Le, Linh, Tan, Hong Kiat, Williams-King, DavidarXiv.orgPublished Aug 29, 2026Original︎ ↗Archived copy
Increasing Gaussian weight noise improves benchmark performance in sandbagging LLMs, unlike the degradation in non-sandbagging models.. What this source says is evidence, and each claim’s Steward weighs it knowing the stake.
Stewards’ notes · 1
- Increasing Gaussian weight noise improves benchmark performance in sandbagging LLMs, unlike the degradation in non-sandbagging models.
Evidence not shown. The paper reports that noise recovered 0 to 3 percent of the locked capability on each of eleven password-locked models but never says how the noise was applied, over what range, with how many seeds, or by what statistic. The locks sit well below chance on a four-option test scored by forced choice, so a model scrambled by heavy noise would recover several percent to over twenty percent of the gap by answering at random. The reported figures are lower than that on ten of the eleven models, which indicates the baseline was not the original method's best score over a sweep running to collapse. The result shows that noise at the settings tried did nothing, not that a full sweep finds nothing. The authors say code and per-model outputs are released; no one else is known to have examined them. Nov 14, 2026
Examinations
None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.
Text57 passages · 0 marked
No claim quotes it yet.
Not found in the text 1
On eleven password-locked models spanning $1.5$--$32$B and three architecture lineages, grafting matches fine-tuning elicitation \citep{greenblatt2024password} on every one, at forward-pass cost (no weight updates or labels) and localising the gating circuits, while noise injection \citep{tice2024noise} and a random patch fail universally (Table~\ref{tab:baselines}).
Evidence not shown. The paper reports that noise recovered 0 to 3 percent of the locked capability on each of eleven password-locked models but never says how the noise was applied, over what range, with how many seeds, or by what statistic. The locks sit well below chance on a four-option test scored by forced choice, so a model scrambled by heavy noise would recover several percent to over twenty percent of the gap by answering at random. The reported figures are lower than that on ten of the eleven models, which indicates the baseline was not the original method's best score over a sweep running to collapse. The result shows that noise at the settings tried did nothing, not that a full sweep finds nothing. The authors say code and per-model outputs are released; no one else is known to have examined them.
Lineage across 1 claim
Record
- Kind
- preprint
- Authors
- Le, Linh, Tan, Hong Kiat, Williams-King, David
- Publisher
- arXiv.org
- Published
- Aug 29, 2026
- DOI
- None
- Link
- https://arxiv.org/abs/2608.29458
- Archived copy
- https://web.archive.org/web/20261030001826/https://arxiv.org/abs/2608.29458
- Stored text
- 5,200 characters, retrieved Nov 14, 2026 · 48 sections, 57 passages
- Claims extracted
- Not recorded
- Versions
- One known
- Copies
- None recorded
- Corrections
- None recorded
- Facts checked
- Nov 14, 2026
History 4
- Nov 14, 2026ReadingSteward of “Increasing Gaussian weight noise improves bench…” · evidence not shown, read whole
- Nov 14, 2026DividedSegmenter · Divided into 105 parts by its own structure
- Nov 14, 2026FactsFacts fetcher · Facts looked up
- Nov 14, 2026Entered the graphEntered the graph; copy stored