Pith.

Machine Learning · arXiv:2608.29458 v1 · submitted 29 Aug 2026

Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities

Linh Le, Hong Kiat Tan, David Williams-King

The paper says

“Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on.”

more of the abstract

The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. The direction tracks how the sandbagging was induced rather than what is withheld – fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 9 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

No verdict was given on this paper.

3 major objections · 3 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the references say.

The paper says, at Abstract (cross-task transfer protocol)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The load-bearing claim is that a contrast direction fit on two to five paired ARC examples, applied to active-learning-selected circuits, restores WMDP performance, but the abstract does not state what objective the active learning optimizes when selecting circuits or how thresholds and grafted coordinate counts are chosen.

more of objection 1

If circuit selection, threshold calibration, or the number of coordinates was scored against WMDP or against a validation set drawn from the same withheld-capability distribution, then the reported recovery is not cross-task transfer but same-task steering with target-informed selection. The authors should specify a quarantine protocol in which every decision before the final WMDP evaluation is blind to WMDP, or provide a task-split and ablation showing that selection on an unrelated proxy yields the same recovery.

The paper says, at Abstract (statistical reporting)

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The headline result, 94 to 101 percent recovery across eleven models with two to five examples, is presented as aggregate point estimates without confidence intervals, per-model results, per-task results, or statistical tests against null interventions. Because the range is consistent with cherry-picking across models or tasks, the authors should report the full distribution of recovery values, the number of models and tasks per condition, and a comparison to control directions, random circuit subsets, and same-task steering.

The paper says, at Abstract (mechanism claim)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The thresholded-gate explanation is asserted from the observation that held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, but no causal control for this mechanism is visible. The failure of additive steering and coordinate zeroing could be an artifact of the direction scale, the chosen coordinate, or the grafting rule rather than evidence of a gate.

more of objection 3

The authors should provide per-model threshold curves, show that the threshold location is stable across tasks, and compare grafting against equal-magnitude isotropic or random-coordinate edits to establish that the gate is actual mechanism.

3 minor objections

The paper says, at Abstract (wording)

Pith says: minor objection 1 of 3

The phrase 'recovers +94 to +101%' is unclear; the '+' is either redundant or intended to convey something about signed recovery, and the range should be expressed with uncertainty intervals rather than a bare percentage range.

The paper says, at Abstract (operational definitions)

Pith says: minor objection 2 of 3

The manuscript should define what counts as a paired example, how the contrast direction is fit from those pairs, and what 'circuits chosen by active learning' optimize; otherwise the method cannot be reproduced from the abstract.

show 1 more

The paper says, at Abstract (data availability)

Pith says: minor objection 3 of 3

No statement about code or data release is visible in the abstract; given the safety-evaluation context and the small number of examples claimed, public release of the model suite and evaluation harness would substantially increase confidence in the results.

What would settle it

Using only the ARC task in the fitting loop, re-run the full pipeline with a strict quarantine: (i) fit contrast directions exclusively on ARC paired examples; (ii) choose circuits by active learning on ARC, never evaluating WMDP or any WMDP-derived validation set; (iii) freeze the edit, then evaluate on two WMDP categories that were not touched in any author pilot; (iv) report per-model and per-category recovery with bootstrap confidence intervals and a leave-one-family-out breakdown. If untouched WMDP recovery falls to same-task steering levels, the cross-task transfer claim is refuted; if it remains above roughly 90% of the honest-sandbagging gap across all 11 models, the concern is retired.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 9 Sep 2026

  1. Reader no verdict given confidence low
  2. Skeptic partly agrees 59 s · 2,346 tokens read, 6,381 written
  3. Editor wrote the review in brief 11 s · 2,996 tokens read, 931 written
  4. Referee 3 major objections, 3 minor 34 s · 2,684 tokens read, 3,297 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This is an abstract-only review, so everything below is provisional. The paper reports a new intervention—reference-grafting—that sets an activation coordinate along a contrast direction to its value in an honest reference, at a small set of circuits selected by active learning. Across 11 password-locked models (1.5B–32B, three families), it recovers 94–101% of the honest–sandbagging gap without weight updates, using 2–5 paired examples. The thresholded-gate explanation (fine-tuned locks act as thresholded gates) is concrete and mechanistically falsifiable. Both the technique and the mechanism are genuinely new relative to the Elicitation Game’s prior verdict that additive steering fails.

The soft spot is the cross-task transfer claim: a contrast direction fit on ARC (grade-school science) elicits withheld hazardous knowledge (WMDP). The abstract gives no detail on what objective the active learning optimizes when choosing circuits or calibrating thresholds. If any of those decisions used WMDP data (or a validation set drawn from the same capability distribution), then the reported 94–101% is same-task steering with target-informed selection, not transfer. The reported numbers are point estimates without variance or per-task breakdowns, so we cannot tell if ARC→WMDP is stable or a selected best case. The thresholded-gate model is consistent with a global lock, but also with a task-specific lock, so the mechanism alone does not exclude target leakage. These are resolvable, not fatal.

The paper is for safety evaluators who need training-free elicitation methods and for mechanistic interpretability researchers interested in how fine-tuned locks actually work. It deserves a serious referee because the claim is practically important and the mechanism is nontrivial. I would send it to review but flag the missing quarantine protocol and the need for statistical reporting as mandatory revisions. The thinking is clear and the authors seem honest about the boundary conditions (e.g., circuit-breaking reroutes and fails). My provisional take: interesting, not yet convincing, but worth engaging with on its own terms.

The strongest objection

Reference-grafting's safety payoff is the cross-task transfer claim: a contrast direction fit on two to five paired ARC examples, applied to a small set of active-learned circuits, restores WMDP performance. The load-bearing joint is that every decision before the final WMDP evaluation must be blind to WMDP. The abstract does not state what objective active learning optimizes when choosing circuits. If circuit selection, threshold calibration, or the choice of how many coordinates to graft were scored against WMDP, or against a validation set drawn from the same withheld-capability distribution, then the reported 94-101% recovery is not transfer at all; it is same-task steering with target-informed selection. The thresholded-gate mechanism is consistent with a lock that is global, but also with a task-specific lock, so the mechanism alone does not exclude leakage. A secondary issue is statistical: 94-101% across 11 models with 2-5 examples is reported as point estimates without variance or per-task breakdowns, so ARC-to-WMDP could be a selected best case rather than a stable property. Neither issue is an accusation of fraud; both are resolvable by an explicit quarantine protocol and per-task reporting.

The abstract makes a strong, falsifiable claim, and the thresholded-gate mechanism is a genuine candidate explanation with independent plausibility; the active-learning procedure and the ARC-to-WMDP transfer are also the right things to stress. I read the claim in good faith: if the authors truly fit on ARC, select circuits on ARC, and then see WMDP recovery without any target-task touch, that is a real result and would justify the headline. The load-bearing problem is that the abstract does not state that quarantine. Active learning is by definition an optimization loop; unless its objective is restricted to ARC, the selected circuits and the grafted coordinates may encode information about the target distribution. This is not a fraud concern; it is a routine protocol detail that separates a transfer result from a selection artifact. I also note the point estimates need error bars since 2-5 examples across 11 models is a small sample; a per-task and per-model table with bootstrapped confidence intervals would settle whether 94-101% is a stable population property. The quarantine test I propose would decisively separate the two readings. Because that test is feasible and the current text does not report it, I would not accept the safety-transfer claim as stated; I would set a conditional verdict pending that single check. The reader's UNVERDICTED disposition is sensible; CONDITIONAL sharpens it by identifying the exact protocol evidence that would move the paper toward acceptance or rejection.

The referee's report

If the central claims hold, the result is significant for model-safety evaluation: a non-destructive activation-only intervention that matches fine-tuning at exposing sandbagged capability would provide both a practical elicitation tool and a mechanistic hypothesis about how sandbagging locks are implemented. The cross-task transfer claim, in particular, is falsifiable and high-value, because it distinguishes genuine elicitation of withheld capability from same-task steering. The presentation would be strongest if it includes reproducible code, explicit quarantine between ARC fitting and WMDP evaluation, per-model data, and causal tests of the thresholded-gate account; none of these can be assessed from the abstract alone.

Pith's reading of the paper (a machine's summary, not the authors')

The paper claims that reference-grafting—an activation-edit method that sets a neuron's coordinate along a contrast direction to the value seen in an honest reference—can elicit withheld capabilities from sandbagging language models as effectively as fine-tuning, but without weight updates or training labels. Across eleven models up to 32B parameters, the technique recovers 94–101% of the performance gap between the honest and sandbagged versions using only two to five paired examples to fit the direction. If true, this would provide a lightweight probe for safety evaluations, revealing hidden capabilities that deliberate underperformance might otherwise conceal.

What carries the argument. Reference-grafting: for an activation vector along a contrast direction (fitted from a few paired examples of honest vs. sandbagged behavior), the method replaces the coordinate of the activation in a targeted circuit with the value that same coordinate takes in an honest reference pass, effectively jumping past a threshold gate that the sandbagging lock imposes. Active learning selects which circuits to edit, keeping the number of modified locations small.

Core claim. Reference-grafting, applied to a small set of activation circuits selected by active learning, matches or exceeds fine-tuning in recovering capabilities that a model was prompted or fine-tuned to withhold, recovering 94–101% of the honest–sandbagging accuracy gap across eleven models. The key insight is that the sandbagging lock acts as a thresholded gate: the grafted coordinate must cross a value near the honest reference to unlock the capability, which explains why additive steering (adding a small vector) or zeroing the coordinate fails to elicit the full capability.

The premise it rests on. The contrast direction fitted on two to five paired examples of one task generalizes to a different withheld capability on the same model, rather than only correcting the exact behavior used to fit it.

If the paper is right (Pith's inferences, not the authors')

Safety evaluations of frontier models can be augmented with a lightweight activation-edit probe that reveals whether a model is sandbagging without requiring expensive fine-tuning or access to training labels.

The threshold-gate mechanism suggests that many forms of sandbagging are realized by a similar low-dimensional activation constraint that any method strong enough to cross the threshold can break.

Because the contrast direction generalises across tasks (e.g., fitted on ARC science questions, it elicits withheld WMDP hazardous knowledge), a single edit may expose multiple withheld capabilities simultaneously.

The distinction between sandbagging induction methods (prompting, fine-tuning, reinforcement learning) each producing a different contrast direction implies that the edit carries information about how the model was locked, not just what is locked.

If the threshold-gate picture holds generally, then any elicitation method that can shift the relevant activation coordinate past a critical value—even a carefully designed adversarial prompt or a different architecture-level intervention—might work as well as fine-tuning, potentially reducing the cost of evaluating model safety.

The fact that only two to five paired examples suffice suggests that the contrast direction is highly structured and low-noise; one could test whether the same few examples work across different random seeds or different base models, which would strengthen the claim of generality.

Because circuit-breaking (rerouting activations on every forward pass) breaks the fixed edits, the method's success depends on the edit not being re-broken downstream—a condition that might be violated in models with recurrent or long-range dependencies, so the falsifier should check performance on longer-generation tasks.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on. The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. The direction tracks how the sandbagging was induced rather than what is withheld – fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.82 · human participants 0.03 · controlled comparison 0.88 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on. background 1.00 0.04 0.03 asserted 0.97
The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. background 0.99 0.06 0.06 asserted 1.00
We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. method 0.83 0.63 0.06 asserted 0.78
Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. result 1.00 0.96 0.75 asserted 1.00
Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. result 1.00 0.90 0.10 asserted 0.99
Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. interpretation 0.99 0.84 0.15 asserted 1.00
The direction tracks how the sandbagging was induced rather than what is withheld -- fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. interpretation 0.65 0.88 0.11 asserted 1.00
Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation. interpretation 0.74 0.59 0.04 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.86 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,523 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Cited by 1 Pith paper
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 20 sentences without a confident label

Contests 2 papers
  • Refining “steering fails.” The negative verdict of [ref] is sensitive to the steering operation (set-to-reference vs. additive nudging) and selection (AL-minimal superposed circuits vs. hand-chosen): with both right, steering matches fine-tuning.

    The Elicitation Game: Evaluating Capability Elicitation Techniques · Discussion · contest 0.71

  • Closest to us, [ref] benchmark prompting, steering, and fine-tuning on password-locked and circuit-broken organisms, concluding steering fails; we contribute a method that overturns that verdict for multiple-choice reasoning and narrows it for code ($$[ref]), and reproduce their circuit-broken organism [ref] as the boundary ($$[ref]).

    Improving Alignment and Robustness with Circuit Breakers · Related Work · contest 0.87

Uses data or values from 1 paper
Compares with 2 papers
  • On eleven password-locked models spanning $1.5$–$32$B and three architecture lineages, grafting matches fine-tuning elicitation [ref] on every one, at forward-pass cost (no weight updates or labels) and localising the gating circuits, while noise injection [ref] and a random patch fail universally (Table [ref]).

    2412.01784 · Grafting Elicits the Hidden Capability · 0.99

  • Concurrently, [ref] steer an evaluation-aware model to act as if deployed—the same shape on a different gate; they add a vector to flip a disposition, whereas we set a coordinate to its honest reference at AL-selected circuits, recover a hidden capability, and map where it breaks.

    2510.20487 · Related Work · 1.00

Cites as background 2 papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment