PaperlayerView exact revision on arXiv

Source-linked AI summary

You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model

Ziyang Luo, Zhongyao Chu, Xinjie He, Youting Wang, Xukui Qin, Runxiong Wu, Yan-Syuan Chen

arXiv:2608.14465v1cs.CLcs.LG

TL;DR

Frozen language models must both recover evidence encoded in their residual stream and abstain when inputs are insufficient, but steering and sufficiency reading interfere in one pass. YOPO learns a label-free reconstruction of the pre-steering residual and reads a fixed sufficiency direction on it, achieving 0.798 three-way accuracy on 1.5B αNLI and beating the two-pass reference across scales and ten backbones.

  • Problem

    Frozen reasoning models under-use encoded evidence and answer confidently when inputs are insufficient, motivating systems that extract supported information while abstaining when necessary.

  • Method

    YOPO learns a label-free reconstruction of the pre-steering residual from steered states, then reads a fixed zero-shot sufficiency direction on the reconstruction in one forward pass.

  • Results

    0.798 three-way accuracy on 1.5B αNLI beats the two-pass reference at every scale and extends across ten backbones spanning six model families.

  • Takeaways & Limitations

    A single forward pass can combine reasoning intervention with calibrated abstention when abstention remains a fixed read and the steering perturbation is learned away.

  • Takeaways & Limitations

    The constructed αNLI and cross-dataset axes contain surface shortcuts and fail as reliable benchmarks, so architectural claims rely on native-label replications.

Abstract

from arXiv · show

A frozen language model on reasoning tasks has two coupled weaknesses: it under-uses evidence its own residual stream already encodes, and it fails to detect when the input is insufficient to answer, so it confabulates. This paper consolidates two research lines that address these on the same residual stream: a conditional steering probe writes the stream at mid-stack layers and recovers reasoning accuracy from a frozen backbone, and a zero-shot sufficiency direction reads the stream and abstains when information is insufficient. Deployed in one forward pass they interfere: the steering write shifts the state the direction reads, costing up to 8 AUROC points of cross-domain transfer on small models; a separate clean pass doubles inference cost. We keep the direction fixed and train a small network to reconstruct the pre-steering residual from the steered one -- mean-squared error on (steered, clean) pairs, no sufficiency labels -- and read the direction on the reconstruction. The resulting system, YOPO (You Only Pass Once), answers, steers, and abstains in one forward pass of a frozen Qwen2.5 backbone (1.5B/3B/7B). End to end, three-way accuracy more than doubles the frozen baseline (0.375->0.798 on 1.5B alphaNLI) and one pass beats the two-pass reference at every scale (0.798/0.830/0.893 vs 0.753/0.790/0.863) and on ten backbones across six model families. We chart the capacity-transfer frontier quantifying the principle that abstention should not be trained in; a source-side audit catches our own alphaNLI construction leaking a surface artifact, so architectural claims are anchored on native-label replications (SQuAD2, RepLiQA, MuSiQue); and on the standard four-domain suite we contribute, to our knowledge, the first answer-or-abstain benchmark, where our gate tops every in-domain dataset and the label-free direction is the only gate family to survive domain transfer.

1 INTRODUCTION

YOPO combines residual-stream steering and zero-shot sufficiency detection in one forward pass of a frozen language model. A label-free reconstruction corrects steering interference, improving answer-or-abstain performance while preserving transfer-oriented sufficiency signals.

  • One-pass interference: 0.962→0.913 in-domain and 0.918→0.836 on cross-task transfer quantify the gate’s degradation when reading the steered rather than clean pass.A separate clean pass restores the gate but doubles inference cost.
  • Capacity–transfer frontier: Task-specific capacity monotonically buys in-domain accuracy while costing transfer, defining a fused-setting capacity–transfer Pareto frontier.The paper quantifies this design law and uses it to motivate the proposed full-capacity gate.
  • Answer-or-abstain benchmark: The first answer-or-abstain benchmark on the standard four-domain suite finds the supervised gate leads in-domain on all four datasets, while the label-free direction transfers best across domains.The label-free direction is fit on the hardest source.
  • Label-free correction: YOPO learns a small map M with label-free MSE to reconstruct clean sufficiency-layer residuals from steered residuals while keeping direction d fixed.Training uses free (steered, clean) pairs obtained by running the extractor with hooks on and off.
  • End-to-end results: 0.798 fused one-pass 3-way accuracy on 1.5B αNLI exceeds the frozen baseline’s 0.375, steering-only’s 0.590, and gate-only’s 0.560.The fused system beats the two-pass reference at every scale: 0.798/0.830/0.893 vs. 0.753/0.790/0.863.

2 RELATED WORK

Prior activation-steering methods use fixed or learned residual-stream directions, whereas this work makes steering input-conditioned and studies its interaction with sufficiency reading in a fused one-pass setting. It repairs the write by reconstructing the unperturbed state from the perturbed one, following a denoising formulation.

  • Activation steering: Activation steering spans contrastive directions, inference-time shifts, representation engineering, task vectors, low-rank rewrites, and conditional gating of fixed vectors.These approaches modify the residual stream with predefined or learned steering directions.
  • Activation steering: The steering probe instead makes the steering direction input-conditioned, distinguishing it from prior fixed-direction approaches.Its per-layer sub-probes read residuals, emit an input-conditioned unit direction and gated magnitude, and write the edit while keeping the backbone frozen.
  • Fused one-pass setting: The fused one-pass setting reads through an active steering write, addressing an interaction neither activation steering nor sufficiency work treated alone.The related-work discussion identifies this interaction as the paper’s added setting.
  • Reconstruction as repair: The proposed repair trains a map like a denoiser to reconstruct the unperturbed state from the perturbed state, using the steering write as corruption.This formulation follows denoising approaches that recover clean representations from corrupted inputs.

3 SETUP: A WRITER AND A READER ON THE RESIDUAL STREAM · 4 THE FUSION PROBLEM: STEERING CONTAMINATES THE READ

Sections 3–4 define a frozen-backbone writer and zero-shot reader on the residual stream, then show that the writer’s intervention contaminates the reader, especially at small scale. Supervised repair improves in-domain performance but follows a frontier that trades transfer for fit, motivating label-free fusion.

  • 3 SETUP: A WRITER AND A READER ON THE RESIDUAL STREAM: The conditional steering writer uses a small MLP at mid-stack layers to read residual states and write a gated directional update while leaving the backbone frozen.The update is r′ = r + a ∥r∥u(r), and the probe trains on answer cross-entropy.
  • 3 SETUP: A WRITER AND A READER ON THE RESIDUAL STREAM: The sufficiency reader forms a direction from sufficient-minus-insufficient hidden-state means and scores the pre-answer projection, with thresholds fit in-domain or re-derived label-free for transfer.The setup includes read-layer rules matched to the construction.
  • 4 THE FUSION PROBLEM: STEERING CONTAMINATES THE READ: 8.2 transfer points are lost at 1.5B and 3.1 at 3B when the fixed direction reads steered rather than clean residuals; the 7B transfer cost is 1.3 points.At 7B, the in-domain effect inverts, with steered residuals becoming a better read substrate; the conflict is concentrated at small scale.
  • 3 SETUP: A WRITER AND A READER ON THE RESIDUAL STREAM: The evaluation reshapes αNLI, HellaSwag, and RepLiQA into full, partially redacted, and replacement conditions for in-domain and zero-shot sufficiency transfer.RepLiQA uses human-written fictional documents absent from the web, ruling out memorization.
  • 3 SETUP: A WRITER AND A READER ON THE RESIDUAL STREAM: AUROC(full vs. replace) measures the sufficiency score on held-out validation items, with reads taken from one steered forward pass unless marked as two-pass.Results are reported both in-domain on αNLI and under zero-shot transfer.
  • 4 THE FUSION PROBLEM: STEERING CONTAMINATES THE READ: Running the reader on the writer’s pass changes the same fixed direction’s input state, isolating contamination as the cost of one-pass fusion.Table 1 compares clean and steered passes at the same read layer.
  • 4 THE FUSION PROBLEM: STEERING CONTAMINATES THE READ: Supervised gate training restores in-domain accuracy but teaches task-specific behavior rather than sufficiency, causing transfer collapse under shift.In the fused setting, added supervised capacity moves along a Pareto frontier: up in-domain and down in transfer.

5 METHOD: A LABEL-FREE CORRECTION OF THE STEERING PERTURBATION

YOPO corrects the steering-induced perturbation by reconstructing the clean residual from the steered residual, then applies a fixed zero-shot sufficiency direction in the same forward pass. The reconstruction is trained without sufficiency labels, learning only the geometry of the steering write while preserving the gate’s task-general character.

  • One-pass architecture: YOPO writes the residual stream, reconstructs the clean state with M, and reads the fixed sufficiency direction d to answer or abstain in one pass.The system uses injection layers {12, 16, 20}, a mid-layer read, and threshold τ for abstention.
  • Label-free correction: M is trained on paired steered and clean residuals, requiring no annotations or sufficiency labels and leaving d fixed.The pairing uses the same input with steering hooks off and on, with one extra training-time forward.
  • Reconstruction variants: Two reconstruction variants are defined: a single-layer low-rank residual corrector and a multi-layer MLP over concatenated steered residuals from the injection span.The single-layer form is M(h) = h + U σ(V h) with rank r=64; the multi-layer variant reads multiple injection-layer residuals.
  • Label-free correction: The method learns the effect of the steering write rather than the sufficiency decision, so the reconstruction objective cannot encode when to abstain on αNLI.Its task-general behavior follows from the probe’s write being a smooth function of the residual it reads.
  • Scope of label-free: “Label-free” excludes sufficiency-label optimization from the read path, but the steering probe remains gradient-trained on answer labels and d is computed from labeled source examples.The label-free designation therefore concerns abstention training, not the entire system’s supervision.

6 RESULTS: THE ONE-PASS PARETO FRONTIER, ACROSS SCALE

YOPO’s one-pass residual correction defines a scale-dependent Pareto frontier: small models trade transfer for in-domain performance, while larger models increasingly avoid contamination. The correction recovers substantial small-model transfer loss, approaches the clean ceiling at 3B, and becomes effectively identity at 7B.

  • 1.5B frontier: 0.836 → 0.888 transfer and 0.913 → 0.944 in-domain: the single-layer correction recovers 63% of 1.5B contamination without sufficiency labels.The multi-layer map reaches 0.959 in-domain, matching the 2-pass clean ceiling, while transfer is 0.869.
  • 1.5B frontier: 0.944 → 0.982 in-domain while transfer falls 0.888 → 0.859 as task-specific capacity increases, quantifying the 1.5B design frontier.No one-pass gate beats the 2-pass clean read’s 0.918 transfer, but the label-free map narrows the gap from 8.2 to 3.0 points.
  • 3B frontier: 0.984 in-domain at 3B exceeds the clean ceiling, with transfer within 0.006 of it: 0.917 versus 0.922.Every 3B variant improves both columns over the floor, flattening the frontier relative to 1.5B.
  • 7B frontier: 0.995 versus 0.985 in-domain at 7B: the steered read beats the clean read, while transfer remains within 0.013 and the learned map is approximately identity.The reconstruction objective produces this no-harm behavior without relying on a hyper-parameter.
  • Scale trend: 0.082, 0.031, and 0.013: the clean−steered transfer gap shrinks at 1.5B, 3B, and 7B, making de-contamination most consequential for small-model deployment.Frozen-model steering gains are +21.7 pp at 1.5B versus +4.96 pp at 7B.
  • Benchmark diagnostics: 0.98–1.00 on constructed RepLiQA/BoolQ transfer axes and 0.56–0.68 on the hard deletion variant show benchmark saturation, not gate superiority.The discriminative constructed axis is therefore HellaSwag; the easy axis detects document replacement, while the hard axis floors every gate, including the 2-pass clean read.

7 THE STEERING PROBE’S BENCHMARK AND THE COMBINED SYSTEM

The section benchmarks the steering probe’s injection geometry and evaluates YOPO as a one-pass system that combines steering with abstention. YOPO combines both components, surpasses standalone and two-pass alternatives in-domain, while transfer depends on steering-probe generalization and gate capacity.

  • Steering probe benchmark: Spread positions outperform last-token injection at every layer count, while layer count matters little and steering gains are larger when the frozen baseline is weak.At 1.5B αNLI with three layers, spread positions yield +20.2 versus +14.7 for last-token injection; the best 7B αNLI cell reaches 81.1.
  • Combined system: YOPO is evaluated using 3-way accuracy on four-condition validation sets, comparing frozen, steering-only, gate-only, two-pass, and one-pass policies with percentile-calibrated quantile-transfer thresholds.Correctness requires answering answerable conditions correctly and abstaining on the no-information condition.
  • Combined system: 0.798 fused accuracy on 1.5B αNLI exceeds the frozen baseline’s 0.375, steering-only’s 0.590, and gate-only’s 0.560, showing that neither component suffices alone.Steering-only never abstains, whereas gate-only forgoes the reasoning gain; their combination beats both everywhere.
  • Combined system: 0.798 vs. 0.753, 0.830 vs. 0.790, and 0.893 vs. 0.863 show YOPO beating the two-pass reference at 1.5B, 3B, and 7B, respectively.The corresponding one-pass gate AUROCs are 0.989, 0.993, and 0.998.
  • Transfer: 0.754 is the best one-pass transfer accuracy measured for the 7B label-free single-layer map, while 1.5B HellaSwag steering-only reaches 0.502 versus a 0.500 baseline.The HellaSwag result indicates that the αNLI-trained single-task steering probe does not transfer its answer gain there, leaving transfer value to the gate.

8 NATIVE LABELS: BEHAVIORAL VS. INTERNAL, AND THE FUSED SYSTEM

Native-label replications show that sufficiency is strongly encoded internally even when answering behavior force-answers, and that read-only abstention transfers better than writing abstention into the answer pathway. YOPO therefore keeps abstention on the read side while repairing steering-induced interference label-free, preserving one-pass answer fidelity.

  • Native-label evidence: 124/125 unanswerable RepLiQA items are separated by the internal read, which catches all 44 cases that behavioral answering force-answers.Behavioral decision accuracy reaches ≈0.89, whereas the replicated read achieves ranking AUROC 0.9988.
  • Write-access ladder: Under distribution shift, the native-label write-access ordering is read > constrained-write > free-write, with free-write producing the highest force-answer rate.In-domain performance follows the opposite ordering, read < constrained write < free write.
  • Fused native-label system: At 7B, fused and clean in-domain performance match at 0.929, while behavioral and probe results replicate the earlier study within 0.008.The replication reports zero-shot judgment 0.855/0.874, probe on SQuAD2 0.859, and cross-task strict/median/ranking 0.834/0.917/0.960.
  • Direction transfer: 19 of 20 off-diagonal direction-transfer cells at 7B reach ≥0.795 ranking AUROC across five native-label datasets.MuSiQue adds multi-hop questions with native answerability labels and balanced 2,000/1,000 splits.
  • Training abstention into the writer: Recalibrating a trained abstention head recovers +6–8 pp everywhere, but its write contributes +7–8 pp in-domain and zero-to-negative points on transfer.The results identify a decision-layer calibration failure rather than a representation failure, while calibrated clean read remains superior.
  • Interference analysis: Under transfer, abstention ranking remains 0.9337 steered versus 0.9338 clean, while the answer head loses 6.4 pp from write and abstain-gradient effects.Only 1.9 pp comes from the write itself; approximately 4.5 pp comes from abstention-class gradient shaping of the shared steering vector.
  • Architectural rule: The read-side abstention architecture reaches 0.797±0.002 under shift, beats every alternative, and gives up nothing in-domain while retaining one-pass steering.The broader trained-abstention system reaches 0.793 ± 0.003 transfer and exceeds the calibrated behavioral read at 0.784, but scale remains important.

9 BEYOND QWEN: TEN BACKBONES, SIX FAMILIES

The identical YOPO pipeline is evaluated without retuning on seven additional backbones spanning five model families, extending the analysis beyond the original Qwen family. Layers are mapped by proportional depth, Qwen hyperparameters are reused verbatim, and each cell uses a single seed.

  • Cross-backbone evaluation: Seven additional backbones span five other model families in a direct re-run of the probe, extraction, direction, gate, and benchmark pipeline.The models are Qwen2.5-0.5B, OLMo-2-1B, TinyLlama-1.1B, StableLM-2-1.6B, SmolLM2-1.7B, Phi-3-mini-3.8B, and Mistral-7B.
  • Cross-backbone evaluation: No hyperparameters are retuned: layers are mapped by proportional depth and Qwen hyperparameters are used verbatim.The evaluation uses a single seed per cell, as reported for Table 5.
  • Cross-backbone evaluation: The cross-family experiment is organized around three reported regularities, including comparison of the one-pass flagship with the two-pass reference.The supplied passage introduces this comparison but truncates its quantitative result.

10 THE STANDARD SUITE: SELF-AUDIT, REPLICATION, AND THE FIRST HYBRID BENCHMARK

This section audits the paper’s constructed benchmarks, independently replicates the sufficiency comparison on a standard four-domain suite, and introduces its first hybrid answer-or-abstain benchmark. The audit exposes construction artifacts, while the hybrid results show strong in-domain gating, label-free transfer, and domain-specific limits from answering capability.

  • Self-audit: 0.908 AUROC at layer 11 exposes a surface artifact in the four-condition αNLI construction, with layer 12 reaching 0.955 before question-context integration.HellaSwag is milder but flagged at 0.793 on layer 11; the source-side test requires no target data.
  • Replication: Layer 19 is the unanimous causal-onset peak across all four domains, excluding early shortcut axes from read-layer selection.The independent protocol selects the read layer by the AUROC peak over layers ℓ≥L0 and calibrates cross-domain thresholds from unlabeled target inputs.
  • Replication: 0.826 → 0.843 → 0.909 mean in-domain judgment accuracy favors increasing write freedom, but cross-domain ordering reverses and the read-only direction exports the gain.LoRA takes all four in-domain domains; across 12 cross-domain arenas, the read-only direction gains +8.6 pp mean in 9 of 12 arenas over the target’s zero-shot spoken judgment.
  • Hybrid benchmark: +18.3 pp on RepLiQA, reaching 98.6% of the oracle gap, is achieved by OR-composing the model’s abstention with the sufficiency gate.Answered-precision rises from 0.459 to 0.636, while adversarial SQuAD2 systems tie at +0.8 pp.
  • Hybrid benchmark: The in-domain flagship read-head is top on all four datasets, while the label-free direction reaches the in-domain bar on SQuAD2 and MuSiQue-C in cross-domain deployment.The cross-domain direction is fit on MuSiQue-Full without target labels and trails only on RepLiQA; OR-composition beats gate-priority everywhere.
  • Answering and judging: 0.588 → 0.555 and 0.535 → 0.516 show that forced chain-of-thought lowers spoken judgment on the MuSiQue variants.Keeping the prefill gate unchanged excludes intermediate reasoning tokens from the judge’s input and unlocks answering only after the sufficiency decision.

11 WHAT FAILED

Several registered alternatives failed to improve robust transfer: supervised reads stayed on or below the frontier, learned layer combinations never beat the naive floor, and standardization helped only 7B. A BCE-trained demap exceeded the clean ceiling at 3B but did not extend to 7B, while the failure diagnosis replicated across native-label data and MuSiQue.

  • Registered failures: 0.846–0.867 transfer at 1.5B: supervised MLPs on steered reads were on or below the frontier despite 0.972–0.979 in-domain performance.This covered both d-anchored and full variants.
  • Registered failures: Never better than the naive floor at any scale: the learned layer-combination read failed to improve performance.A per-feature-standardized variant repaired 7B (0.983/0.967) but degraded 1.5B/3B transfer.
  • Scale-local result: 0.9324 ± 0.0023 over five seeds: the BCE-trained demap exceeded the clean ceiling on 3B transfer, but did not extend to 7B.Every seed was above 0.922, so the authors report this as scale-local rather than a general claim.
  • Replication: Three datasets, one diagnosis: the failure replicated on native-label data and MuSiQue across all six fusion analyses.The registered recipe is reported with its failure, while the identity-anchored single-layer map is identified as the scale-robust variant.

12 DISCUSSION

The evaluation supports YOPO through broad tests spanning injection geometry, capacity–transfer tradeoffs, end-to-end one-pass performance, native-label replication, and construction auditing. Deployment should follow a budget- and scale-conditioned recipe: use the clean read when a second partial pass is affordable, otherwise choose between the flagship and label-free map based on model scale and task shift.

  • What this evaluation establishes: The evaluation spans injection geometry, capacity–transfer frontiers, end-to-end one-pass comparisons, native-label studies, and a source-side construction audit.These components form the paper’s broad evaluation of frozen-model reasoning with abstention.
  • A budget-conditioned recipe: When a second partial forward is affordable, the clean read remains the transfer-optimal gate.This is the recommended choice under a larger inference budget.
  • A budget-conditioned recipe: Under a hard one-pass budget, the flagship multi-layer de-contamination plus boost leads in-domain at every scale and transfers best at 3B and above.The passage identifies this flagship as the default one-pass recipe.
  • A budget-conditioned recipe: At 1.5B under worst-case task shift, the label-free map is preferred because the boost’s transfer score falls from 0.888 to 0.859.The passage treats this transfer cost as consequential for small-scale deployment.

13 CONCLUSION

YOPO shows that a frozen model can combine reasoning steering and sufficiency reading in one forward pass when abstention remains a calibrated read and steering perturbations are reconstructed away. The work also emphasizes reproducibility, public-data use, frozen reversible interventions, and reduced confident confabulation.

  • Architecture: A single forward pass can host both a reasoning writer and sufficiency reader when abstention remains a calibrated read and steering perturbations are removed label-free.Training abstention into the system collapses outright at small scale and does not beat calibrated reading under shift at 7B, whereas reconstructing the write perturbation transfers.
  • Reproducibility: Every reported number is recomputable from committed per-item logs, scored JSON artifacts, and the Appendix C scripts.Residual extractions and trained probes are also regenerable; large binaries are omitted only because of repository size.
  • Data and ethics: The study uses only public benchmark datasets and open-weight backbones, with no human subjects, personal data, or annotation labor.The listed benchmarks include αNLI, HellaSwag, PIQA, BoolQ, SQuAD 2.0, RepLiQA, and MuSiQue, used under their published licenses via canonical releases.
  • Deployment: Inference-time interventions remain reversible because the backbone models are completely frozen and edits can be removed by deleting a forward hook.The interventions are hot-swappable and do not modify model weights.
  • Safety: Abstention when context is insufficient is intended to reduce confident confabulation, identified as a primary failure mode of deployed assistants.The system’s stated purpose is harm-reducing rather than weight-updating or model retraining.

A THE STEERING INJECTION-GEOMETRY GRID · B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS · C REPRODUCIBILITY

The appendix tests steering geometry, decomposes hybrid transfer failure, evaluates repairs, and documents recomputable protocols. It finds decision calibration and reconstruction—not simply representation quality—are central to reliable abstention under transfer.

  • A THE STEERING INJECTION-GEOMETRY GRID: The steering grid evaluates injection layer count, position scheme, and αmax on αNLI, HellaSwag, and PIQA against frozen baseline B0.Each probe uses 5k training items and 1,532 held-out evaluation items.
  • B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS: +6–8 pp transfer recovery comes from quantile-calibrating the hybrid’s p(“0”) score, indicating most transfer loss is a decision-layer failure.Examples include MuSiQue→HellaSwag 0.682 →0.744 and MuSiQue→αNLI 0.7325 →0.810, with ranking AUROC 0.918–0.970.
  • B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS: +7–8 pp in-domain gains from writing contrast with zero-to-negative transfer contributions, where calibrated clean reads reach 0.784 on HellaSwag and 0.815 on αNLI.The write improves MuSiQue from 0.790 to 0.864 and αNLI from 0.8125 to 0.885, but transfer is 0.744 versus 0.784 and 0.810 versus 0.815.
  • B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS: 0.624 on HellaSwag is achieved by the raw read–write split, while steering reduces gate AUROC 0.917 →0.858 and answer accuracy 0.820 →0.719.The split fails against targets above 0.744 and 0.784 because the answer-writer perturbation swamps the cross-domain read without reconstruction.
  • B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS: 0.811 is the best answer-head result among write-side repairs, but abstain reweighting, KL anchoring, and PCGrad all trade answer fidelity against gate quality under shift.PCGrad yields gate 0.891 and worst net score 0.716, while λ=0.25/0.5 gives answer heads 0.789/0.781 and gates 0.900/0.920.
  • B TRAINING THE ABSTENTION IN: DECOMPOSITION AND REPAIRS: 0.782/0.786 transfer from stronger suppression buys no more than λ=0.3 while reducing in-domain performance to 0.860/0.8575.A hinge objective reaches 0.859 ± 0.018 in-domain and 0.781 ± 0.013 transfer across three seeds.
  • C REPRODUCIBILITY: All reported numbers are recomputable from committed per-item logs and scripts covering the steering grid, fused run, extraction, analysis, cross-matrix, and replication workflows.The steering grid uses 5k items, seed 0, learning rate 10−4, three epochs, batch four, and 1,532 held-out items.
  • C REPRODUCIBILITY: The probe protocol uses an 80/20 example-ID split, seed 7, held-out selection of layer and threshold before evaluation, and median cross-dataset calibration.The cross-matrix is pure post-processing of stored extractions, with a layer sweep over the shared capture set and one JSON per scale.
Loading 2608.14465v1…