acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

A frozen language model under-uses the evidence in its hidden states and answers confidently when the context contains no answer. We fix both on the residual stream without changing any weight. For accuracy, a small trained network (the probe) adds an input-dependent vector to the residual at several middle layers; a removable hook, it raises Qwen2.5-1.5B on αNLI (abductive reasoning) from 51% to 73% accuracy. For abstention, a fixed vector (the direction), the mean hidden state of contexts that contain the answer minus that of contexts that do not, scores the hidden state; a low score abstains. On four QA datasets with human answerability labels (Qwen2.5-7B), moving the direction to a dataset it was not fit on costs under 3 percentage points of decision accuracy (from 0.83 in domain) on average and still beats asking the model whether its context suffices; the same judgment trained as a classifier into the probe's activation edit or into LoRA weights loses 16 and 22 points and ends below asking. We set out to put both functions on a single forward pass, and found that the probe's edit changes the very hidden state the direction scores, degrading it. A small map — trained without labels, only to reconstruct the unedited hidden state from the edited one — removes 63% of that interference on 1.5B. Refit on the hidden-state pairs of LoFiT or LoReFT, the same recipe repairs those methods too: the steering module is replaceable. The resulting system, YOPO, answers and abstains in one forward pass. On αNLI it raises three-way accuracy (correct answer when the context suffices, abstention when it does not) from 0.38 to about 0.75 on 1.5B, and stays within 0.02 of a two-pass reference, which scores a separate probe-off pass, on 1.5B, 3B and 7B, at half the cost. We also observe improvements on the four QA datasets, where answerability is human-labeled and answers are generated as free-form text. Near-parity with the two-pass reference holds across nine backbones spanning six model families without backbone-specific hyperparameter tuning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
activation steering, representation engineering, abstention, answerability, parameter-efficient fine-tuning, cross-domain transfer, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_U9suOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an approach for the unification of probe-based steering (which the authors call "writing"), and abstention-classification (which the authors call "reading"), into a single forward pass. Validation is done over several steering and classification probes.

Critical strengths
  1. The paper is exceptionally well-written
  2. The experimental procedure is exceptionally clear and, for the (insufficient) baselines and related work outlined, thorough
Critical weaknesses

1.The probing and abstention-classification literature is vast, and this work fails to reference much of it. Alternative probes cited/tested against in this work are old. The direction methodology exists already in the literature but the authors of this work fail to attribute it to any previous work at its introduction in the paper. Much of the steering literature, and even classification probing for MCQA tasks, SFT is generally compared against in the experimental results to validate the motivation of probing over more LLM training. 2. I think this work lacks any major novelty: linear abstention probes and MLP-based steering probes already exist in the form presented in this work. Perhaps the only novel contribution is the proposal of a single-forward-pass methodology which does both steering and classification (writing and reading). But, the impact of this is fairly low; it halves the inference-time cost. Also, being a simple log-linear model to recover the pre-edit representation, it’s not a particularly novel solution. With that said, it is surprising that the re-reconstruction map, which is shallow and presumably much less expressive than the hidden LLM layers whose effects it must predict, actually works. This is not explored in the work, though. 3. While the writing is good and the paper is clear, It is missing any real introductory motivation. While the contributions are laid out clearly, the authors make no effort to demonstrate their importance. Why is a writer abstraction useful? I don’t think it is, ``steering’’ is good enough. Why do we care about cross-domain generalization in this context? The one paragraph at the start of the section is insufficient, and also makes claims which go unsupported even by reference to existing work. Rather than presenting results for a methodology which has not yet been introduced in the introduction, I think the space should be allocated for a more careful explanation of the problem, the setting, and why the existing approaches are unsatisfactory. Even in the related work section at the end, the explanation of why these existing methods fall short is not given.

I'm rating this paper, as it stands currently, as a weak accept primarily based on the quality of the writing and the thoroughness of the experiments (with a couple exceptions). If the authors can (a) address the somewhat lacking grounding in the probing literature, (b) convince me that I have missed some novel aspect of their work, and/or (c) motivate their work more clearly, I will be open to raising my rating.

Questions for authors
  1. To my knowledge, the abductive-NLI dataset always has only two options, meaning random accuracy is 50% assuming a balanced dataset. The fact that Qwen 0.5B gets 50% accuracy without using a probe writer makes me wonder if, as often happens in LLM-based classification accuracy experiments, the LLM answers correctly but the output is not parsed properly. It seems hard to believe that a model that seems to have no ability to predict a task can improve based solely on a probe injection. Perhaps it is moreso the fact that the probe gets to train directly on the task. In that case, does the probe injection do better than SFT directly on the model’s weights?
  2. In table 1, are you training the probe and the LoRA adapter still as steering probes or as classification probes? That is, are you supervising directly on the probe output or are you still supervising on the eventual output through all the frozen LLM layers? I think the literature typically treats abstention classification as the former, but in l. 235 you say that neither LoRA or the probe architecture is changed.
  3. In your experiments showing that only the static direction reader generalizes cross-domain, did you replicate the median-based threshold calibration for the other methods as well, or only for the direction-based one?
2 comments
AuthorsOpenReview ↗

Response to Reviewer U9su (Part 1/2): W1–W3

We sincerely thank the reviewer for the thoughtful review and constructive suggestions. Your questions helped us clarify why this problem matters, what our work contributes, and how to present the paper’s overall argument more coherently.

Weaknesses

W1. Grounding in the probing and abstention literature

Reply: Thank you for highlighting these gaps. We have revised the introduction and contributions (§1) and expanded the related work (§6) to improve clarity and logical flow:

  1. Attribution and contributions. The introduction and §2 now explicitly credit Marks & Tegmark (2024) and Lavi et al. (2025) for the direction method. Contribution (i) now describes matched comparisons of existing methods; our subsequent contributions concern interference, reconstruction-based repair and single-pass integration.
  2. Related work and baselines. We expanded §6 and added comparisons with a hidden-state logistic-regression classifier (Table 1), confidence scores, self-consistency and conformal calibration (App. F), alongside the existing LoRA and activation-steering baselines. These results support our conclusion that the direction effectively reads answerability information already present in the model’s hidden states and maintains stable performance across the evaluated domains.
  3. Comparison with SFT. We compare supervised LoRA fine-tuning under matched data and epoch budgets; LoRA outperforms our activation-steering module (§3). LoRA’s stronger performance also benefits YOPO: replacing our module with LoRA and refitting the map improves end-to-end accuracy in our comparison (§5.3).

W2. Novelty and why a shallow map works

Reply: Thank you for highlighting these points. We have clarified three aspects:

  1. Novelty. The individual components are established. Our contribution is a systematic study of interference between answer-improving interventions and hidden-state answerability readouts, and its repair without additional answerability labels. YOPO supports interchangeable interventions—including our activation-steering module, LoFiT, LoReFT and LoRA—with the reconstruction map refitted for each. This enables improved answer generation and answerability assessment within a single forward pass.
  2. Single-pass inference. YOPO achieves performance close to the two-pass reference while avoiding a separate judgment pass—an important computational saving for large-scale deployment.
  3. Why a shallow map works. We have expanded §5.2, highlighted in blue, with a mechanistic explanation and supporting diagnostics (App. K). Our nonlinear residual MLP is trained by MSE on paired edited and unedited states. For abstention, it only needs to restore the original scalar answerability score, rather than invert the full intervention. The 7B diagnostics show that the perturbation is concentrated in a low-dimensional subspace and that the affected component of the pre-edit score remains approximately linearly predictable from less-affected coordinates. Under this recoverability condition, a small correction can suffice.

W3. Motivation, terminology and cross-domain evaluation

Reply: Thank you for helping us clarify the motivation. We have revised the introduction (§1) to explain the problem, the gap in prior work and the supporting references more clearly.

  1. Problem and motivation. Reliable context-based QA should answer accurately when evidence is sufficient and abstain otherwise. Prior work shows that hidden states can contain information supporting both capabilities that is not fully reflected in model outputs. Steering and answerability probing typically address these capabilities separately; our goal is to integrate them into a single system to improve QA reliability through better answers and answerability assessment. In combining them, we find that answer-improving interventions can disrupt the answerability signal. We therefore introduce a reconstruction map, trained without additional answerability labels, to repair this interference. YOPO combines answer improvement and answerability assessment within a single forward pass and supports interchangeable answer-improving components, avoiding a separate judgment pass and reducing inference cost at scale.
  2. Why “writer”. We agree that “steering” can describe these interventions. We use “writer” only to distinguish the component that modifies the residual stream from the answerability readout.
  3. Why cross-domain evaluation. Our goal is to uncover and use the model’s existing, generalizable ability to assess context sufficiency. If the readout captures this capability, it should remain useful beyond the source dataset; cross-domain stability is therefore an important test of this claim.
AuthorsOpenReview ↗

Response to Reviewer U9su (Part 2/2): Q1–Q3

Questions

Q1. Chance-level accuracy, parsing and comparison with SFT

Reply: Thanks for the question.

  1. Parsing. Accuracy is computed directly from the two option-token logits; no generated text is parsed.
  2. Supervision and hidden information. Our module receives supervised task training. A separate 1.5B diagnostic finds 70.6% accuracy from a supervised linear readout of frozen hidden states versus 50.8% from the model’s own logits, supporting the presence of underused task information (App. N).
  3. SFT comparison. Please see our response to W1, point 3, in Part 1.

Q2. Where is the supervision applied in Table 1?

Reply: Thanks for the question. The probe remains an activation-steering module, and LoRA remains a weight-adaptation module. Their architectures are unchanged; only the training target changes to context-sufficiency judgments. Both are supervised through the LLM’s final answer-token logits, with gradients flowing through the frozen backbone. We also added a logistic-regression classifier trained directly on frozen hidden states; this comparison further supports the direction’s more stable cross-domain performance (Table 1).

Q3. Was median calibration applied to all methods?

Reply: Thank you for highlighting this. Originally, only the direction used median calibration; we have repeated the comparison using the same calibration rule for all methods.

At 7B, calibration reduces cross-domain accuracy drops from 16.4 to 7.5 points for the activation-steering baseline and from 22.2 to 12.4 for LoRA; the direction drops 2.3 points. Although the gap narrows, matched calibration and threshold-free AUROC continue to support the direction’s more stable cross-domain performance across all four backbones (Table 1; App. J).

Thank you again for the time and care you have devoted to reviewing our work. Your feedback has been valuable in improving the paper, and we would be happy to address any remaining questions or concerns.

Reviewer_XQvNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks how a frozen language model can both answer more accurately and abstain when its context is insufficient, in a single forward pass and without changing any backbone weight. The proposed system, YOPO, combines four components. Probe module, small MLPS, adds an input-dependent vector to the residual stream at three middle layers and is trained with answer cross-entropy. A fixed “direction” scores whether the context suffices; a label-free low-rank “map”, trained with an MSE loss to reconstruct the unedited hidden state from the edited one, removes the probe’s interference with the direction’s readout, and a threshold on the resulting score decides whether to abstain or return the same pass’s answer. The authors claim that (1) a fixed direction transfers across datasets better than a sufficiency judgment trained into the probe or LoRA weights (§4), (2) on smaller models the probe’s edit lowers the direction’s AUROC and the map recovers much of the loss, including for other writers such as LoFiT and LoReFT, and (3) YOPO matches a two-pass reference within about 0.02 on a controlled αNLI benchmark and within 0.03 on most cells of four human-labelled QA datasets, at half the inference cost.

Critical strengths
  1. The map (the network that projects and modifies I assume) is quite promising and is trained only with a reconstruction loss and needs no sufficiency labels. The same recipe, refit per writer, also works for LoFiT, LoReFT, and LoRA (Table 2). On 1.5B these writers hurt the direction’s AUROC even more than the probe does (in-domain AUROC as low as 0.730–0.811), and the map recovers 77% to all of the loss. The pipeline is also run on ten backbones from six families without retuning.

  2. The details are sufficient. Appendix I documents ten failed variants (supervised MLPs, multi-layer maps, domain-adversarial training, etc.) and states which pre-registered hypotheses were not supported. Appendix H discloses a padding-position implementation issue and re-evaluates with the fix. The post-hoc change of the map initialization is flagged, with both settings reported. This lets readers judge how much to trust each result.

Critical weaknesses

Weakness of parameter scaling ability. Is it only applied in small language models? HellaSwag AUROC loss shrinks from 0.083 at 1.5B to 0.016 at 7B, and at 7B the probe actually raises the in-domain AUROC (0.985 → 0.995). In Table 3 (αNLI, 7B), one pass without the map scores 0.880, higher than YOPO with the map (0.858) and the two-pass reference (0.863), and Appendix I(x) admits the map discards the probe-on advantage there. Since experiments stop at about 8B, I have no clues about how it perform in larger models like 14B or MoE structures models. But regarding of computational resource limitation, I might take less account of this part.

Unfair comparison in §4. The paper says trained judgments (probe, LoRA) fail to transfer across datasets while the fixed direction transfers well, but the two are not treated equally. The direction gets a threshold tuned on the target data (the median of its scores, which assumes a 1:1 class ratio), while the probe and LoRA get no adaptation at all. Appendix E suggests that much of the trained methods’ drop is just a miscalibrated threshold, not a failure to transfer. Please give all methods the same thresholding and retest with unbalanced class ratios.

However overall I think these weakness are quite minor. I will consider strong acceptance depending on author's reponse.

Questions for authors

Does the paper show that the hidden states already contain enough to produce the answer? The motivation is that the frozen model “under-uses the evidence in its hidden states”, but this is never directly tested. I would like to see further statement about it.

Is it possible to provide some samples / example outputs for abstaining and answering that is being tested in benchmark?

1 comment
AuthorsOpenReview ↗

Response to Reviewer XQvN

We sincerely thank the reviewer for the constructive feedback and for recognising the promise of the reconstruction map and our experimental transparency. Your suggestions motivated additional evaluations on larger models. These additions have strengthened our evidence and helped clarify our conclusions. We address each point below.

Weaknesses

W1. Does the method extend beyond small models?

Reply: Thanks for your question. We added three complementary evaluations that support the effectiveness of our method on larger models:

  1. Repair on larger models. We evaluate 14B and 32B models. The map recovers 82% of lost HellaSwag AUROC at 14B with LoFiT and 85% at 32B with our activation-steering module (Apps. K, M).
  2. Stronger-intervention stress tests. To test repair under increased interference, we scale intervention strength up to 5×. The map recovers 64–85% of lost HellaSwag AUROC at 32B (App. K).
  3. Full-system scaling. To evaluate the complete pipeline at larger scales, we also test six additional backbones spanning 8B–32B (App. M).

We have also expanded §5.2 to explain why a small reconstruction map can suffice even for larger models.

W2. The direction receives target-side threshold adaptation while the trained judgments do not. All methods should receive the same thresholding, including under unbalanced class proportions.

Reply: Thank you for highlighting this point. We have repeated the comparison using the same calibration rule for all methods.

Matched calibration reduces the 7B cross-domain accuracy drop from 16.4 to 7.5 percentage points for the activation-steering baseline and from 22.2 to 12.4 for LoRA. Although this narrows the gap, our conclusion remains supported: the direction drops only 2.3 points at 7B and shows more stable cross-domain performance across all four backbones (Table 1). Threshold-free AUROC further supports this finding (App. J, Table 32).

We additionally test unanswerable proportions of 10%, 30%, 50%, 70% and 90%. Results across these proportions also support the direction’s more stable cross-domain performance (App. J.3).

Questions

Q1. Does the paper show that the hidden states already contain enough to produce the answer? The motivation is that the frozen model “under-uses the evidence in its hidden states”, but this is never directly tested.

Reply: Thanks for the question.

  1. Writer. We added a small diagnostic (App. N): a supervised linear readout of the frozen 1.5B model’s hidden states achieves 70.6% accuracy, versus 50.8% from its own answer logits on the same αNLI development items.
  2. Reader. Our QA comparisons show that hidden-state readouts can identify insufficient contexts even when the model still attempts to answer (§4).

Both findings support useful information being present in hidden states but not fully reflected in the model’s output.

Q2. Is it possible to provide some samples / example outputs for abstaining and answering that is being tested in benchmark?

Reply: Thanks for the question. App. L provides the following examples:

Writer (αNLI). “You are a careful abductive reasoner. Given two observations, pick the hypothesis that better explains how observation 1 leads to observation 2. Answer with just ‘1’ or ‘2’.” The frozen model answers 2, the probe-on model answers 1, and the gold is 1.

Reader (RepLiQA). “Judge whether the passage contains enough information to answer the question. Answer 1 (enough information) or 2 (not enough information).” The model’s own answer is “enough”, the direction says “not enough”, and the gold is “not enough”.

End to end (SQuAD 2.0, one pass). “Answer concisely based only on the passage. If the passage does not contain enough information to answer, reply ‘unanswerable’.” The model generates an answer, but YOPO outputs “unanswerable,” matching the gold label.

We hope that the additional experiments and clarifications address your concerns about model scale, calibration and the information available in frozen hidden states. Thank you again for the thoughtful suggestions that helped us strengthen the evaluation and better establish the scope of our findings. If any concerns remain, we would welcome the opportunity to clarify them or provide further experimental details.

Reviewer_yENrOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper analyzes the presence and recoverability of the abstention signal (decision to abstain from answering the question in the case of high uncertainty or so) in the residual stream of attention model. The authors introduce YOPO - a system that combines probe, hidden-layer steering, and recovering map in a single framework to make abstain decision and produce answer in a single forward pass through the model.

Critical strengths

S1. Quite an interesting idea to target improvement of accuracy and abstention capabilities in a single go.

Critical weaknesses

W1. Text requires substantial polishing; it feels very unnatural and delivers its conclusions in an overlycomplicated way; certain efforts are required to grasp the main points of the narrative.

  • The list of the main contributions is overloaded with minor details, significance of which isn't very clear (e.g., mention of LoFiT/LoReFT/LoRA in (i)).
  • "Writer" abstraction is tautological, and it is unclear how "any edit to the residual stream" can be considered as a main contribution of this paper.
  • Tables 2 and 4 have certain results in bold, but the reason behind this highlighting is not specified, and it is confusing - those are not the highest nor the lowest results.
  • In the conclusion, the paper advises to "decide abstention by reading the hidden state, never by training it" - what does it mean? There are no experiments with the model fine-tuning in the paper.

W2. The evaluation is rather limited, and the results are inconclusive:

  • Claim regarding "nine backbones spanning six model families" is somewhat overstated; while technically it is true, the main results are conducted on a much smaller selection of models. The experiment that covers the whole range of models is performed on αNLI\alpha NLI only, which has certain concerns regarding it (below) . All experiments are performed on models up to 7B parameters; no larger models were used. Considered models are rather dated; no Thinking-tuned models were covered. The proposed methods also can not be applied to the closed-weight models.

  • Most of the covered datasets are rather dated (2018-2022); some of them have issues, which the authors themselves acknowledge, for instance α\alphaNLI, where low-level dataset features provide a shortcut for the abstention task. Coverage of modern abstention benchmarks (e.g., AbstentionBench or others) would much benefit the weight of the results.

  • Baselines are very few in number. The paper compares against a prompted baseline and against trained judgments, but not against existing abstention methods: self-consistency, conformal prediction, or other hidden-state probes (although such works are cited in the related works section).

  • No confidence intervals reported for the results of the main experiments. Given the fact that the proposed method often shows very marginal gains over the baselines, this undermines the weight of the conclusions.

Questions for authors

See Weaknesses.

3 comments

W1

Text...polishing...

Reply: Thanks. We have revised the paper throughout, with all changes highlighted in blue. The key updates are:

  1. We have revised the whole introduction to explain the problem and motivation more directly following the logic structure: problem background, gap in the existing work, motivation for our method, and our contribution. In this way, the logic would be easy to understand.
  2. We have expanded §5.2 to explain the reconstruction map, a central component of YOPO, and why it works.

The main contributions overloaded...

Reply: Thanks. Mention of LoFiT/LoReFT/LoRA: These established methods serve as alternative activation steering components within our system YOPO. We mention these methods to show that YOPO supports different steering and adaptation approaches. Our Contribution Significance:

  1. Identification of intervention–readout interference: We systematically characterize how answer-improving interventions can degrade hidden-state answerability readouts on the same forward pass.
  2. Label-free reconstruction-based repair: A small map recovers lost readout performance using paired hidden states, without additional answerability labels.
  3. YOPO: answering and abstaining in one pass: The system approaches two-pass performance and supports interchangeable interventions, including LoFiT, LoReFT and LoRA.

"Writer" abstraction is tautological...

Reply: Thanks. Our activation steering method, LoFiT, LoReFT and LoRA differ in implementation but all modify the residual stream to influence generation. We use “writer” as a common term for this role; The term itself is descriptive and is not claimed as a contribution. Contribution see above.

Tables 2 and 4...

Reply: Thanks. The caption of Table 2 did not explain the bold formatting. It marked the after-map AUROC of our activation steering method. Since it did not encode a comparison, we have removed it. In Table 4, bold highlights single-pass YOPO results within 0.02 of or above the two-pass reference (steered answer generation and answerability in separate forward passes), as defined in the caption.

The paper advises...

Reply: Thanks. This statement refers to our experimental finding in §4: a fixed direction directly reads answerability information already present in hidden states and remains relatively stable across domains, while trained judgment baselines—activation-steering modules, LoRA adapters (trained via supervised fine-tuning) and logistic-regression classifiers—degrade substantially. We have also removed the categorical “never” wording from the conclusion.

W2.

nine backbones...

Reply: Thanks. We added experiments on six additional backbones spanning 8B–32B: Llama-3.1-8B-Instruct, Qwen3-8B, Phi-4, DeepSeek-R1-Distill-Qwen-14B, Qwen2.5-14B-Instruct and Qwen2.5-32B-Instruct (App. M). The two reasoning-capable checkpoints, Qwen3-8B and R1-Distill-Qwen-14B, were evaluated with thinking disabled. These results extend the evidence for YOPO to larger and newer backbones.

The experiment...

Reply: Thanks for your question.

  1. Our methods are also verified in other datasets such as SQuAD 2.0, RepLiQA and two MuSiQue variants not only in the [αNLI].
  2. The context replacement in αNLI may introduce surface cues, which we discuss as a potential limitation. Our findings are also supported by four QA settings with native answerability labels, so they do not rely solely on this synthetic construction (§5.5). We agree that this limits YOPO’s applicability: it requires access to and control over hidden states, so it cannot be used through standard black-box APIs. It is intended for deployments with internal model access.

Most...

Reply: Thanks. We consider these two datasets SQuAD 2.0 and MuSiQue which are in AbstentionBench. We additionally evaluate RepLiQA (2024), which was designed to reduce overlap with pretraining data.

Baselines are very few in number...

Reply: Thanks. We have added self-consistency (App. F), a hidden-state logistic-regression probe (Table 1), and conformal calibration (App. F, Table 20). These results continue to support our conclusions: the direction generally achieves higher answer-or-abstain accuracy than confidence-based readers and shows smaller cross-domain degradation than trained readers.

No confidence intervals reported...

Reply: Thank you for raising this concern.

  1. Confidence intervals. App. F now reports 95% paired-bootstrap intervals for all Table 4 comparisons; 24 of the 31 improvements over the frozen model have intervals entirely above zero.
  2. Interpreting the gains. We aim to improve over the frozen model and match the two-pass reference in a single pass. Small differences from the two-pass reference reflect our intended goal: matching its performance while eliminating the separate judgment pass.

Thank you for providing response to my review. I have also read other reviews and replies to them.

I appreciate the new revision of the paper; it addresses my concerns regarding the writing - now it is much more concise, clear, and human-accessible. New experiments provide the necessary extension of the evidential support.

I still have some concerns regarding proper integration of the new material into the structure of existing paper and distribution of the key elements between main text and appendices. Also, the interaction of the proposed method with modern Thinking-tuned models is still not explored (experiments on Qwen3 were done with thinking mode disabled).

Overall, I acknowledge the substantial improvements of the paper, and I will raise my score for it (to Weak Accept).

Thank you for taking the time to reassess our paper in light of the revisions and additional experimental evidence. Your feedback helped us make the writing clearer and strengthen the experimental support for our claims.

During the discussion period, we prioritized completing and reporting the additional experiments. We see room to integrate the new material more smoothly into the paper and to improve how key content is distributed between the main text and the appendices.

For consistency across models, we evaluated both reasoning-tuned models in the direct-answer setting, with thinking disabled. With more time, we would have liked to compare YOPO’s performance on these models with and without a generated reasoning trace.

Thank you again for your thoughtful feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.