acceptodds
Under review as a conference paper at ICLR 2027

Position-Sensitive Concept Acquisition in Text-to-Image Models

Abstract

Pretrained text-to-image models possess rich visual priors, yet reliable control over quantitative attributes can remain difficult to acquire through supervised fine-tuning. We study how the position of a target condition affects this acquisition process. Across models, concepts, and evaluation settings, paired fine-tuning with identical image supervision and matched training budgets yields markedly different control capabilities when the same target field is placed at the beginning or end of the prompt. The effect depends on both the model and the concept. Using angle control with Qwen-Image as a representative case, we find evidence linking the acquisition gap to information propagation in the frozen causal text encoder and the use of target information during training. A simple Prefix fine-tuning scheme repeats the target condition at the beginning and improves control over standard end-position training when the prefix is retained at inference. We further introduce a pre-adaptation probe of relative preference for correct over counterfactual conditions. Cross-model comparisons and controlled pretraining experiments illustrate three characteristic acquisition patterns under finite budgets: weak control at both positions, a pronounced positional gap, and stronger control at both positions. These patterns are associated with initial support and learning progress, rather than fixed support thresholds. Acquired control may remain format-dependent even as positional gaps narrow.

Primary area:
generative models
Keywords:
text-to-image generation, concept acquisition, causal text encoder, supervised fine-tuning, prompt position, multimodal generative models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A2MrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether the placement of a quantitative target condition in the prompt affects the model's ability to learn that control during T2I fine-tuning. Across Qwen-Image, Z-Image, and FLUX.2 Klein, the authors compare matched Front- and Back-position training for angle, ellipse, and progress-bar control. The paper further analyzes Qwen-Image angle control, finding that an early condition propagates more broadly through the frozen causal text encoder and produces a stronger generator response. They propose Prefix fine-tuning, which repeats the target value at the beginning, and studies concept-learning dynamics in a controlled synthetic setting by varying prior concept exposure and fine-tuning progress.

Critical strengths
  • While inference-time positional bias is well known, the observation that target-field position can substantially change the acquisition of quantitative control under matched supervision is interesting.

  • By evaluating Front-trained models with Back prompts, they separate training-position effects from purely inference-time position effects.

Critical weaknesses
  1. Front and Back prompts are explicitly not token-identical reorderings; grammatical adjustments are made between formats. Thus, the observed differences can reflect position together with wording, syntax, or tokenization and do not strictly isolate the effect of position. Additionally, evaluating only the extreme start and end points is limiting. A stronger test would use a fixed structured template with identical tokens and move the target field systematically across intermediate positions to demonstrate a clear, graded relationship between position and control.

  2. The paper shows that early conditions spread through more contextual representations and that Prefix both increases this spread and improves control. However, Prefix simultaneously changes position, repetition, and prompt length, which the paper itself acknowledges. A critical missing experiment is testing a model with a bidirectional encoder (like T5), where later tokens can influence earlier ones. If the Front/Back performance gap persists in a bidirectional model, the causal-attention hypothesis would be invalid. Relatedly, Ask Twice, Look Twice [1] independently shows strong ordering effects in causal VLMs and resolves them by repeating the question before and after the image, which is particularly relevant to the paper's Prefix intervention and should be discussed.

  3. The main Front/Back results in Table 1 do not appear to report variability across independent fine-tuning runs. Given the small training datasets, reporting results across different seeds, which the authors correctly did for their synthetic and Prefix studies, is essential to prove the stability of the central claims, I also could not find sufficient details on how the generated angle, ellipse dimensions, and progress-bar proportions are extracted from images. Since the main normalized-MAE metric entirely depends on these extractions, the reliability of the underlying measurement tool must be established.

[1] Abhinandan et al., Ask Twice, Look Twice: Prompt Echoing Resolves the Question-First Paradox in Vision-Language Models. eXCV Workshop at ECCV, 2026.

Questions for authors
  1. Could the authors move the target condition through several intermediate token positions, rather than only Front/Back, while keeping the prompt tokens otherwise identical? Does acquired control decrease systematically with later placement, and does this track contextual propagation?
  2. Does the effect reproduce with a bidirectional text encoder such as T5? This seems like a direct test of the proposed causal-attention mechanism.
  3. Could the authors report Front/Back results over multiple independent fine-tuning seeds for the main Qwen-Image experiments?
  4. How exactly are generated angles, ellipse dimensions, and progress-bar proportions measured, and what is the error/reliability of these evaluators on images with known ground-truth attributes?
  5. Since Front-trained models degrade when tested in Back format and Prefix-trained models degrade when the prefix is removed, would it be more precise to describe the learned capability as format-conditioned quantitative control rather than position-invariant concept acquisition? The paper does acknowledge this limitation, but I think the distinction is important for interpreting the result.
1 comment

R3-1. Does control weaken across positions without rewriting? We added a five-position experiment using five unchanged clauses: "Use black lines on white." "Mark and label {a_label}, {b_label}, and {c_label}." "Place {b_label} at the vertex and {a_label} to its right." "Show a small arc at {b_label}." "Set ∠{a_label}{b_label}{c_label} to {angle_degrees}°." The measurement clause occupies M1-M5 (earliest to latest); the other four clauses are randomly ordered. We train a model per position and evaluate on unseen angles. This fixed-clause setting does not use the original held-out-template split.

Matched train/test positionScore
M10.5042
M20.1542
M30.2542
M40.1992
M50.0454

Overall, later placement leads to poorer acquisition even with clause wording held fixed. Thus, the original grammatical rewriting is insufficient to explain the difference. In the original pre-SFT Front/Back probe at encoder layer 28, context spread is 0.038934/0.010405 (3.74x) and coverage is 50.36%/15.60%. These are two-position diagnostics, separate from the five-position experiment. The separate training-time attention-blocking experiment in R2-1 directly intervenes on propagation and provides complementary acquisition evidence.

R3-2. Does the effect persist with a bidirectional encoder? We added PixArt-Σ, which uses a bidirectional T5 encoder. Matched-format scores are:

TaskSplitFFBB
AngleSame/In0.02490.1609
AngleSame/Out0.04890.0742
AngleDifferent/In-0.1465-0.1366
AngleDifferent/Out-0.1311-0.1279
EllipseSame/In0.08460.0276
EllipseSame/Out-0.03680.0002
EllipseDifferent/In-0.1541-0.1946
EllipseDifferent/Out-0.1964-0.1823

PixArt-Σ does not reproduce Qwen's strong, consistently directed Front advantage. Angle BB exceeds FF throughout; ellipse ordering varies. Unseen-value control remains weak, so small gaps do not establish successful position-invariant control. This architectural comparison complements the within-Qwen intervention.

R3-3. How does Prefix relate to Ask Twice, Look Twice? Question echoing in Ask Twice, Look Twice is closely related to Prefix, but the research questions differ. Abhinandan et al. study input ordering at inference for a fixed VLM. We study how target-condition placement during SFT affects quantitative control learned by a text-to-image model; Prefix is used in both training and inference. We also recognize the shared use of repetition and information-flow interventions, and will discuss this methodological connection explicitly. We developed Prefix independently and will publicly release the corresponding checkpoints and original timestamped training logs to document when the experiments were conducted. We will cite and discuss this work in the final version.

R3-4. Are the main Qwen results robust across independent SFT seeds? For Qwen-Image-2512, we use training seeds 42/43/44 and generation seeds 1/2, averaging generation seeds within each run before computing training-run means and standard deviations. The three-seed robustness results are:

TaskBB (mean ± SD)FF (mean ± SD)
Angle-0.2346 ± 0.12500.3765 ± 0.0624
Ellipse-0.2707 ± 0.04510.4907 ± 0.0804
Progress0.4920 ± 0.02620.5825 ± 0.0361

The mean FF-BB gaps are 0.6111, 0.7614, and 0.0905: the main task-dependent pattern remains after averaging independent training runs. We will report training variability separately from probe uncertainty. Multi-seed robustness results cover Qwen only; no corresponding claim is made for Z-Image or FLUX.2.

R3-5. How are attributes extracted and evaluator reliability checked? R1-2 details rule-based geometric extraction, Hough-based angle measurement, and constant-baseline scoring for unmeasurable outputs. Manual visual inspection confirmed consistency with human judgments. The remaining rule-based measurement procedures will be detailed in the final version.

R3-6. Is this format-conditioned control rather than position-invariant acquisition? We agree that the results do not establish position-invariant acquisition. Acquisition here means learning to control generated attributes according to requested values; unseen-value generalization and transfer across prompt positions are distinct. We will describe the demonstrated capability as format-conditioned quantitative control. Retaining the prefix at inference is part of Prefix; removal exposes its transfer limitation.

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

The paper examines how the position of a condition within a prompt affects the degree of control a text-to-image model acquires over that condition during finetuning. The main contribution is demonstrating that, given identical training images and training budgets, placing the condition at the beginning or the end of a prompt leads to clear differences in control, depending on the specific model and task. The authors evaluate this using three models across tasks involving angles, ellipses, and progress bars. As an explanation, the article provides insights into how information propagates and is used within the causal text encoder. They further show that repeating the condition as a prefix can improve control, given that this same format is kept during inference. Finally, by applying a probe before fine-tuning, as well as with synthetic experiments, the authors link these positional effects to the model’s prior knowledge and the progression of the learning process, though they do not establish a universal rule for predicting these outcomes.

Critical strengths
  1. The phenomenon is interesting and potentially relevant for the adaptation of text-to-image models. Comparisons between training and evaluation positions allow for an analysis that goes beyond just studying the effect of changing a prompt during inference.

  2. The work seeks to understand the phenomenon from several perspectives. The synthetic environment allows for the control of prior visual exposure and the tracking of learning. The selective perturbations described in Appendix B.3 are particularly interesting and useful, as they keep the model fixed while examining which aspects of the condition it utilizes. The authors also acknowledge that greater sensitivity or discrimination does not guarantee better control over the generated outputs.

  3. The paper is clear and engaging to read. Each experiment builds on the findings of the previous one and poses a new question, which keeps the reader interested. As I read, I found myself wanting to see whether the next result would explain the difference between the positions. Although some questions remain open, this progression is well executed.

Critical weaknesses
  1. The mechanistic analysis, despite interesting, does not establish why Front performs better than Back. While showing that an initial condition influences a greater number of tokens confirms an expected property of a causal encoder, it does not demonstrate that this propagation results in better control. Moreover, since the prefix simultaneously alters position, repetition, and length, it is not possible to attribute the improvement uniquely to any one of these mechanisms. Finally, the prior score fails to provide clarity, as similar values ​​are associated with disparate control outcomes, there is a lack of uncertainty metrics, and FLUX.2 is excluded from the analysis without justification.

  2. The synthetic experiment in 5.3 shows that Front can learn faster than Back, but it does not establish the significance of this difference. One would expect the gap will be small when both models fail or both achieve high performance, and with sufficient SFT, this gap tends to disappear. What remains to be measured are the practical effects of this. For instance, how many steps or examples Front saves in order to reach the same level of control. Without that type of comparison, the experiment describes a transient difference in learning, but it does not demonstrate a practically relevant advantage.

  3. Furthermore, there are two aspects regarding the presentation that would help to better evaluate the contribution. First, the selective perturbations described in Appendix B.3 should be moved to the main body of the article, as they constitute some of the most informative evidence regarding the application of the condition. Second, the literature review also requires a little more in-depth comparison with the work of Ma et al. and Merchant et al., in order to clarify what was already established and what this study adds.

  4. As an additional minor point, it would be helpful to add to the experimental description the prompt templates, actual examples of training and generation, and the procedure used to measure the attributes. Since the Front and Back conditions involve grammatical adjustments, providing the actual texts would allow for a better assessment of what changes beyond just the position.

Overall, I believe the studied phenomenon is worthy of attention. However, key details and controls are missing to support the proposed interpretation.

Questions for authors
  1. It would be good to have at least a few examples of the prompts used, the training images, and samples of the model’s output, including the associated measurements, to allow for a more comprehensive evaluation of the tasks and results. Could you include the full templates and these examples, and also explain how you measure angles, lengths, and proportions, as well as how you handle cases where an image cannot be measured?

  2. Can you perform a small comparison experiment, by placing the text block containing the attribute at the beginning, in the middle, and at the end, without making any other changes to the wording? This can be used to show how the control evolves from the pre-fine-tuning stage to a higher budget, together with the corresponding loss. This would also help distinguish between the effects of the formulation, the learning rate, and any persistent differences. Including the middle position would allow for a better characterization of the phenomenon than a comparison limited to just the two extremes.

  3. Why did you choose preference between correct and incorrect conditions as a measure of prior knowledge, rather than using generation control prior to fine-tuning, or considering both measures? A model might be able to distinguish between conditions without being capable of generating the requested attribute. Comparing these measures would help determine whether the gap between the Front and Back conditions stems from information already represented within the model, or from a control capability that the model already knew how to employ.

  4. Why is FLUX.2 not included in Table 4? Given its weak control in the tasks evaluated, its prior score would be useful for examining the scope of the explanation.

I would be willing to reconsider my assessment if the clarifications and some additional, targeted evidence allow for a better distinction between the various explanations to the phenomenon. I do not necessarily expect performance improvements or confirmation of the hypothesis. Results that help define the scope of the phenomenon, for example, by showing that the observed differences disappear when controlling for the wording or by extending the training, could also strengthen the contribution.

1 comment

FF/BB match training/test positions.

R2-1. Does propagation causally improve acquisition? We compare two SFT groups on the same five-template Front angle set in Qwen. In the text encoder, Angle blocks later tokens from attending to the angle-value tokens. Control instead blocks their attention to "texture, quantity," in the prepended system instruction ("Describe the image by detailing ... texture, quantity, ..."). Both masks set selected pre-softmax logits to -infinity, leaving prompt text/order unchanged.

Training targetBlocking at test?Score
AngleNo0.0645
AngleYes0.0575
ControlNo0.4505
ControlYes0.4095

At test, Yes retains the training mask; all rows have 45/45 valid detections. Angle-trained control remains weak even with normal encoding, while Control training retains stronger control. This supports a causal role for target-information propagation; see R1-4 for the suffix control.

R2-2. Can selective perturbations be moved to the main text? We agree and will move the selective-perturbation analysis from Appendix B.3 to the main text.

R2-3. Can you provide prompts, images, and measurement details? Original training pair: Front: "Draw a {degree}° angle ∠{point1}{vertex}{point2} with vertex {vertex}. Let ray {vertex}{point1} extend horizontally to the right, and use ray {vertex}{point2} as the other arm." Back: "Start at {vertex} with ray {vertex}{point1} extending horizontally to the right. Add ray {vertex}{point2} to form ∠{point1}{vertex}{point2} with a measure of {degree}°." In/Out: seen/unseen templates; Same/Different: seen/unseen values. See R1-2 for evaluation and R1-3 for settings. Full templates and annotated images will be provided in the final version.

R2-4. What does this add beyond Ma et al. and Merchant et al.? Ma et al. establish positional bias in causal prompt encoders; Merchant et al. show benefits of consistent versus shuffled caption structure. Both Front and Back formats are consistent. We study target placement during SFT under matched supervision/budgets, cross-format transfer, and dependence on initial support and learning progress.

R2-5. What happens at intermediate positions with wording fixed? R3-1 shows weaker acquisition overall with later placement under fixed wording; see R2-6 for duration.

R2-6. What practical advantage remains with longer training? Toy convergence does not imply that extra SFT removes the gap in practical models. At α=0.002, averaging both test positions, Front/Back reach 71.4%/62.0% at step 25. Their first reported checkpoints above 70% are steps 25 and 50, respectively. These 25 steps already provide 6,400 exposures over only 18 target combinations and equal 0.25% of 10,000 pretraining steps. At the same 25-step budget, Front leads by 9.4 percentage points, providing better learned control. In Qwen angle Different/Out, the original duration comparison is:

EpochsFFBB
40.1130-0.1609
80.3339-0.1986
160.1409-0.0803

Back remains below baseline at 16 epochs. The narrower gap also reflects Front degradation, not just Back improvement.

R2-7. Why use a discriminative prior instead of generation control? We use the prior probe because, before task SFT, base models often fail to draw the target shape, preventing reliable direct measurement of its quantitative attributes. Rule-based detection (R1-2) gives pre-SFT non-detection rates of 93.21%/88.75%/80.44% for angle/ellipse/progress in Qwen/Z-Image/FLUX.2, confirmed visually. The prediction-error probe measures discrimination, not generation control. P is a pairwise win rate (chance 50%). For the eight Qwen tasks, the prior probe uses 8 target values and 5 images/value: 40 ground-truth images/task, 320 total at 1024×1024. Each image has Front/Back prompts; counterfactuals change only the target value, keeping the image and other text fixed. Five new Qwen tasks use two SFT seeds, with BB/FF on Different/Out only due to compute limits; the original three use three-seed means. Multi-seed results cover Qwen only.

TaskBBFFPrior P
Arrow direction1179.60%
Radial spokes0.84720.819468.30%
Gaussian width-1.42010.198260.40%
Exponential rate-0.31250.359358.90%
Progress0.49200.582557.10%
Angle-0.23460.376554.60%
Ellipse-0.27070.490753.20%
Spiral spacing-0.0166-0.755548.10%

Overall, the trend is: weak support, difficulty at both positions; intermediate support, larger positional gaps; strong support, control at both. Task details will be added in the final version.

R2-8. Why is FLUX.2 absent from the prior comparison? We will add FLUX.2 Klein 4B to Table 4:

TaskPFFBB
Angle0.5429-0.2812-0.4070
Ellipse0.60360.0764-0.1469
Progress0.4929-0.6236-0.5164

This also fits our prior-support hypothesis.

Reviewer_VP5rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the position-sensitive usage of quantitative control in text-to-image generation model. More detailed is that try to place the same target condition in different order, may cause different outcomes under matched fine-tuning supervision setting. The experiments have covered angle, ellipse axis length difference and progress bar control. And test on qwen-image, z-image, and flux.2 with frozen text encoders an trainable image genrators. The paper provides evidence linking these differences to information propagation in a causal text encoder and proposes Prefix fine-tuning, which repeats the target condition at the prompt’s beginning. This improves Qwen-Image angle control over the back baseline when retained at inference, but does not consistently outperform front training. Finally, a pre-adaptation discrimination probe and controlled synthetic-shape experiments examine how initial support and learning progress relate to positional gaps, showing that such gaps can emerge and subsequently narrow during training.

Critical strengths

The experiments for validating the consequence of training-time sensitivity through comparisons seem meaningful. By using the same image supervision setting, the fromt/back experiments show that condition placement can substantially affect acquired quantitative control. Cross-format evaluation further demonstrates that moving the condition forward only at inference does not recover the performance of front-position training. This makes the observation relevant to adaptation practices, beyond conventional inference-time prompt engineering. The controlled synthetic experiments strengthen the investigation of learning dynamics. By varying target-shape exposure while fixing the architecture and pairing fine-tuning runs across data, noise, and optimization settings, the paper shows that positional gaps can emerge and later narrow during learning. Repetition across three fine-tuning seeds provides supporting evidence within this environment. This contributes a useful perspective that position sensitivity can depend on learning progress rather than being a fixed property of a model or concept.

Critical weaknesses

The main empirical results lack sufficient detail and uncertainty analysis to establish robustness. The paper does not adequately specify the large-model adaptation configurations or the procedures for extracting quantitative attributes from generated images, including how invalid outputs are handled. The main comparison also lacks uncertainty estimates across independent fine-tuning runs, and no actual generated-image comparisons are provided. Because the central contribution rests on measured differences in generation control, these omissions make it difficult to assess reproducibility and distinguish reliable learning effects from training variability or evaluation artifacts.

The mechanistic explanation and intervention do not sufficiently isolate the source of the improvement. Front/Back prompts involve grammatical changes beyond strictly rearranging identical tokens, while Prefix fine-tuning simultaneously changes condition placement, repetition, and prompt length. The propagation probes therefore provide suggestive evidence rather than establishing that broader information propagation causes better acquisition. Moreover, Prefix remains dependent on retaining the prefix at inference and underperforms direct Front training on unseen target values. These limitations constrain both the explanatory contribution and the methodological advance beyond the observed phenomenon.

Questions for authors

Evaluation validity: How are angles, ellipse axis-length differences, and progress-bar proportions extracted from generated images? How are malformed or unmeasurable outputs scored, and has the evaluator been validated against manual measurements? Please provide representative paired Front/Back outputs to clarify what the reported score differences correspond to visually.

Reproducibility and robustness: What are the exact adaptation configurations and train/test set sizes for each large-model experiment, including trainable parameters, optimizer settings, and independent training seeds? Do the principal Front/Back gaps remain consistent across independent fine-tuning runs? The reported probe confidence intervals do not establish this training-level robustness.

Evidence for the mechanism: Do you have controlled comparisons that separate condition placement from grammatical changes, repetition, and prompt length. For example, identical-field reordering and equally long prompts with repeated suffix conditions? If these controls are unavailable, what evidence distinguishes the proposed propagation explanation from these alternatives?

1 comment

F/B denote Front/Back; FF and BB match training and test positions.

R1-1. Does the position effect persist without rewriting? We added a fixed-clause experiment covering five target positions without rewriting any clause. Overall, later placement leads to poorer acquisition even with clause wording unchanged. Full settings and matched-position results are in R3-1.

R1-2. How are attributes measured and invalid outputs handled? We convert images to grayscale, threshold foreground/background, remove noise using task-specific tolerances, and extract geometric quantities with rule-based detectors. For angles, a Hough transform detects two rays; their intersection and directions determine the angle. Ellipse axis differences and progress-bar proportions likewise use rule-based geometric measurements; details will be provided in the final version. Through manual visual inspection, we confirmed consistency between evaluator outputs and human judgments. Unmeasurable outputs remain in evaluation and receive the evaluation set's best constant prediction, the baseline in Eq. (1). Post-SFT invalid rates are 0% for angles/progress bars across all three models, and 0.53%/0.53%/0% for ellipses with Qwen-Image/Z-Image/FLUX.2 Klein. Together, the visual checks and low extraction-failure rates support the reliability of our evaluation pipeline for these tasks.

R1-3. What are the adaptation settings, data sizes, and seed variability? Only LoRA weights in the generation backbone are trained; base weights and text encoders are frozen. Shared settings: 8 epochs, one dataset repeat, gradient accumulation 1, BF16, gradient checkpointing, LoRA alpha=rank, AdamW with weight decay 0.01, betas (0.9,0.999), epsilon 1e-8, and ConstantLR.

ModelLRBatch/deviceRankLoRA parameters
Qwen-Image-25122e-4496707,788,800
Z-Image1e-413279,380,480*
FLUX.2 Klein Base 4B1e-413248,168,960

With the scripts' default 8 processes, global batches are 32/8/8. Qwen's count is calculated from local layer dimensions; *Z-Image's is estimated from its default architecture, not verified against a historical checkpoint. Full target-module lists will be provided in the final version. Qwen adapters cover attention projections and selected MLP/modulation layers in the backbone's image/text streams. Z-Image targets to_q, to_k, to_v, to_out.0, w1, w2, and w3. FLUX targets dual-stream attention projections, selected linear layers, and output projections in 20 single-stream blocks. Each task uses 1,000 training pairs. Angle/ellipse/progress evaluation manifests contain 360/380/480 prompt records before generation-seed replication. The three-seed Qwen summary preserves the task-dependent pattern; means and standard deviations are reported in R3-4, separately from probe uncertainty. Multi-seed robustness results cover Qwen only.

R1-4. Can a repeated suffix explain the Prefix gains? On Qwen-Image-2512, we retain the Prefix settings but move the additional target text (e.g., "20°.") to the end. The added content is unchanged; its position differs. We test the suffix-trained model with the suffix removed or retained:

SplitSuffix removedSuffix retainedPrefix retained
Same/In-0.1703-0.17380.5211
Same/Out-0.1100-0.12350.5396
Different/In-0.0731-0.08100.2184
Different/Out-0.0126-0.03640.2369

The Prefix column is the existing mean over SFT seeds 42 and 44, with the prefix retained; it is separate from the three-seed main-task summary. Suffix repetition does not reproduce the Prefix gains; retaining it improves none of the four scores. Repetition alone is therefore insufficient to explain the benefit: placement matters.

R1-5. Prefix remains format-dependent and trails Front on unseen values. We acknowledge that Prefix underperforms direct Front training on unseen target values. It serves two purposes: improving the Back baseline while preserving the original description and adding a target prefix, and providing a diagnostic intervention into how target-information presentation affects acquisition. The prefix is part of the method during both SFT and inference; its removal exposes format dependence. We will make these limitations explicit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.