acceptodds
Under review as a conference paper at ICLR 2027

Task-Identifiability-Aware Action Supervision for Offline In-Context RL

Abstract

In-context reinforcement learning (ICRL) enables policies to adapt to unseen tasks through accumulated interaction context without test-time parameter updates. However, existing approaches typically apply uniform action supervision across context prefixes, overlooking differences in how reliably each prefix identifies the underlying task. A short or ambiguous prefix may remain consistent with multiple tasks that require different actions for the same query state, such that the target action cannot always be uniquely inferred from the available context. To address this mismatch, we propose Tiara (Task-Identifiability-Aware Representation and Action Learning), an offline ICRL framework that makes task inference explicit and uses empirical task-identification reliability to guide action learning. Tiara employs a causal context encoder, independent of the query state, to construct prefix-wise task representations supervised by training-task identities. The correctness of auxiliary task identification serves as an empirical proxy for prefix-level task identifiability and is used to reweight action supervision, assigning greater weight to correctly identified prefixes while retaining a smaller positive weight for ambiguous ones. Experiments across diverse multi-task offline ICRL benchmarks show that Tiara improves generalization to unseen tasks over strong baselines across most settings.

Primary area:
reinforcement learning
Keywords:
in-context reinforcement learning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jZq8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Tiara, an offline in-context reinforcement learning framework that addresses the mismatch between varying task identifiability across context prefixes and uniform action supervision in DPT-style methods. Tiara introduces a causal context encoder with auxiliary task-ID supervision and uses task-classification correctness as a prefix-level reliability signal to reweight action prediction losses. The task classifier and reweighting are used only during training, while deployment adapts to unseen tasks solely through interaction context. Experiments across navigation, locomotion, and manipulation benchmarks show improvements over several offline ICRL baselines in most settings.

Critical strengths
  • Motivation is intuitive and targets a concrete mismatch in DPT-style offline ICRL: different prefixes contain different amounts of task information but receive uniform action supervision.
  • The method is simple and cheap. The per-sample normalization preserves the total action-loss scale, and the weighting scheme reduces to uniform supervision when α=1\alpha=1 , making it a clean drop-in modification. The task classifier and reweighting are train-only, so deployment requires no task IDs and adds no inference cost beyond the encoder–decoder split.
  • The evaluation covers discrete navigation, locomotion, and manipulation, plus a higher-dimensional Humanoid-Dir task, against a reasonably broad set of baselines.
Critical weaknesses
  • The central reliability signal in Eq. (9) is based on binary task-classification correctness. However, correctness is not equivalent to identifiability: an intrinsically ambiguous prefix may still yield a coincidentally correct argmax, while a nearly correct but slightly misclassified prediction is treated as fully unreliable. This conflates the information content of the prefix with the current state of the learned classifier. The paper would benefit from comparing binary correctness against confidence-, margin-, or entropy-based alternatives and from directly analyzing how these signals evolve with prefix length.
  • Task identifiability does not necessarily imply action ambiguity. Multiple tasks that remain indistinguishable under the same prefix may nevertheless prescribe the same optimal action at the query state. In such cases, the task is ambiguous but the action supervision is still fully valid. Since the method reweights an action-prediction loss, the paper should justify why task-ID uncertainty is the appropriate surrogate for action ambiguity and ideally demonstrate a strong empirical relationship between the two.
  • The paper motivates Tiara by arguing that ambiguous prefixes should not be strongly supervised toward a unique task-dependent action. However, Eq. (5) applies exact task-ID cross-entropy supervision uniformly to every prefix, including those that may be intrinsically non-identifiable. This appears partially at odds with the motivation and may encourage artificial task separation from insufficient evidence.
  • The current ablations do not fully isolate the contributions of the decoupled architecture, auxiliary task supervision, and reliability-guided weighting. In particular, the “w/o Encoder” variant changes both the model architecture and the weighting mechanism, making it difficult to attribute the observed gains. Architecture-matched controls with uniform, position-based, and reliability-based weighting would provide a cleaner evaluation of the proposed mechanism.
Questions for authors
  • What fraction of prefixes have ρ=0\rho=0 over training, and what fraction of samples contain both correct and incorrect prefixes such that Eq. (11) actually produces non-uniform weights? Reporting these statistics by training step and prefix length would clarify how active the proposed mechanism is in practice.
  • Tiara assumes that a prefix that cannot reliably identify the task should receive weaker action supervision. However, task ambiguity does not necessarily imply action ambiguity, since multiple indistinguishable tasks may still require the same action at the query state. Can the authors empirically verify that prefixes marked as unreliable indeed exhibit higher target-action disagreement?
  • How does binary correctness compare with continuous reliability signals such as the probability assigned to the true task, classification margin, or predictive entropy under the same architecture?
  • Is the task classifier relying on behavior-policy artifacts rather than task evidence? Could the authors evaluate task identification on contexts collected by a shared or different behavior policy, or report task-ID accuracy using only (s,a)(s,a) versus the full (s,a,s′,r)(s,a,s',r) ?
  • Could the authors provide architecture-matched ablations comparing uniform, position-based, confidence/entropy-based, and binary reliability weighting, together with a capacity-matched DPT? Sensitivity analyses for α\alpha and λ\lambda would also be helpful.
4 comments

We thank the reviewer for the detailed review. Changes in the revised PDF are marked in blue. We number the Critical Weaknesses W1–W4 and the Questions Q1–Q5 in review order.

W1, Q3: Binary correctness conflates prefix information with classifier state; how does it compare with continuous signals?

A: Thank you for this insightful observation. This is intended: the decoder can only use task information already extracted by the encoder, so what matters is whether the current representation identifies the task. In Sec. 5.3 (Table 2), a frozen classifier (same initialization and budget, but not following the encoder) loses about 40% of the gain on Reach and all of it on Walker-Param, and binary correctness outperforms margin, entropy and even the true-task probability, which is gentler on near-misses, in both environments. For a coincidentally correct argmax on an ambiguous prefix, Proposition 1 (Sec. 3.3) shows that the learned target remains a posterior mixture.

W2, Q2: Do prefixes marked as unreliable show higher target-action disagreement?

A: Thank you for this suggestion. Yes (App. G). On Walker-Param, we approximate each task's target action at the query state with a task-conditioned behavior-cloning network and compare misidentified and correct prefixes within the same task and position (three seeds). Misidentified prefixes show 3–8× higher disagreement among the candidate tasks' target actions at prefix lengths 1–5 (AUROC 0.76–0.85), and higher action error. Fewer than 6% of them are assigned a task with a close target action; by Proposition 1, such prefixes only lose gradient share, without biasing the target.

W3: Is applying task-ID cross-entropy to every prefix (Eq. 5), including non-identifiable ones, at odds with the motivation?

A: Thank you for raising this point. We respectfully disagree. Cross-entropy is a proper scoring rule minimized by the task posterior, which is not one-hot on ambiguous prefixes, so it does not force commitment. The action head instead makes a point prediction, which is where uniform supervision is problematic. Empirically, the classifier is calibrated on short prefixes (App. G).

W4, Q5: Architecture-matched ablations, capacity-matched DPT, and sensitivity to α\alpha, λ\lambda.

A: Thank you for these suggestions. "w/o Encoder" changes both architecture and weighting by necessity, not from incomplete decoupling: ρ\rho comes from the encoder's classifier, so without the encoder there is no identification outcome to weight by, and a position schedule is the closest substitute. To separate the two factors:

  • Same architecture (Table 2): all weighting variants keep TIARA’s encoder. Binary correctness is best: TIARA exceeds uniform weights by 123.4 (Reach) and 81.6 (Walker-Param), and the other signals by 20.5–47.8 and 73.7–93.5.
  • Position weights: in the decoder-only model, they improve DPT by 70.0 and 75.1, yet TIARA stays 91.7 and 47.5 higher.
  • Capacity (App. F): a DPT with more parameters than TIARA (1.52× vs. 1.47×) stays below it. TIARA and "w/o Weight" have identical parameter counts, so their gap is due to the weighting alone.
  • Sensitivity (App. F): every α<1\alpha<1 clearly beats uniform weighting, and every tested λ\lambda beats DPT.

Q1: What fraction of prefixes have ρ=1\rho=1 over training, and how often does Eq. (11) actually produce non-uniform weights?

A: Thank you for this question. App. G (Fig. 9, Table 14) reports both by training step and prefix length. The fraction of prefixes with ρ=1\rho=1 grows with prefix length and over training, and samples containing both correct and incorrect prefixes remain common throughout training, so Eq. (11) is active in practice. As identification improves, the misidentified prefixes concentrate at the start of the context: the weighting acts most strongly early in training, on the short prefixes that the representation cannot yet identify. Removing it (α=1\alpha=1) lowers the return substantially (App. F).

Q4: Does the classifier rely on behavior-policy artifacts rather than task evidence?

A: Thanks for this important question. Behavior-policy confounding of task inference is a known issue in context-based RL, since the (s,a,s′,r)(s,a,s',r) pairs in a context reflect both the task and the policy that collected them, and dedicated methods have been developed specifically to address it. This issue is orthogonal to our focus: TIARA does not aim to learn a policy-invariant task representation, but uses the classifier only during training to decide how strongly each prefix supervises the action. Our (s,a)(s,a) versus full-transition comparison (App. G) shows that (s,a)(s,a) alone carries some task information, but the environment’s feedback contributes clearly more, so the signal mainly reflects task evidence. The classifier is discarded at test time, where TIARA acts on its own contexts and achieves the gains in Table 1.

Thank you again for the helpful comments.

Reviewer jZq8: are you convinced by the rebuttal, particularly how classifier correctness is related to identifiability?

The rebuttal addresses most of my main concerns, particularly through the new analyses on action disagreement, alternative reliability signals, architecture-matched ablations, and sensitivity studies. These additions substantially strengthen the empirical support for the proposed weighting mechanism. I still have some reservations about the limited novelty and the possibility of behavior-policy confounding, which is not fully ruled out by the current experiments. Overall, however, I am more convinced than before and would be inclined to raise my score from 2 to 3.

We sincerely thank you for the careful reconsideration and the willingness to raise the score. We are glad that the new analyses addressed the main concerns, and we greatly appreciate the constructive suggestions, which have substantially strengthened the paper.

On behavior-policy confounding, we followed your original suggestion and evaluated the classifier on contexts collected by another task’s behavior policy (App. G). The effect depends on the environment: on Reach, the classifier rarely follows the behavior policy, whereas on Walker-Param it often does, and there a different policy also changes the dynamics observed in the context. We agree that TIARA does not remove this confounding, as it is not designed to; since each training task’s data come from its own behavior policy, this does not affect how ρ\rho guides action supervision during training.

Thank you again for your time and the constructive discussion.

Reviewer_hHx9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies offline in-context reinforcement learning (ICRL), focusing on a limitation: all context prefixes are typically given equal action-supervision weight. The authors proposed TIARA (Task-Identifiability-Aware Representation and Action Learning). Experiments are conducted on multi-task environments and experimental results show that TIARA improves generalization to unseen tasks over strong baselines across most settings.

Critical strengths

The paper is well written and well organized. The motivation is easy to understand, the method is presented clearly.

The method is conceptually simple and practical. The task-identification classifier and reliability weights are used only at training time.

The empirical evaluation is reasonably broad. The main experiments span environments with different sources of task variation.

Critical weaknesses

The novelty appears relatively limited. The main idea consists of adding an auxiliary task-classification objective and using classification correctness to reweight supervised action losses.

An incorrect task-ID prediction may not imply that the action target is ambiguous. This distinction seems important given that “task identifiability” is central in the proposed work.

The gains are not uniformly strong from Table 1.

Questions for authors

Have the authors considered a soft reliability score rather than binary correctness? For example, task-related entropy, confidence margin could distinguish highly certain and marginally correct classifications. How does such weighting compare with Eq. (10)?

Do the results change when the context and query samples are forced to be disjoint? The current loader allows a query transition to also appear in the context. An experiment excluding this possibility would help rule out any benefit from context/query overlap.

2 comments

We sincerely thank the reviewer for the careful reading and the concrete suggestions. All changes in the revised PDF are marked in blue. We number the Critical Weaknesses W1–W3 and the Questions Q1–Q2 in review order.

W1: The novelty appears limited: the method adds an auxiliary task-classification objective and uses classification correctness to reweight action losses.

A: Thank you for this comment. We respectfully disagree that the method reduces to these two components. Learning a task representation with an auxiliary classification loss is a standard building block in context-based meta-RL. Our contribution is to use each prefix's identification outcome to set its action-supervision weight, which targets a concrete mismatch in DPT-style offline ICRL, as Reviewer jZq8 also notes. This step accounts for most of the improvement: with the same encoder and auxiliary loss but uniform weights, TIARA loses 76% (Reach) and 67% (Walker-Param) of its gain over DPT (Sec. 5.3, Table 2). Proposition 1 (Sec. 3.3) further shows that correctness weights change the learned target, whereas task-independent weights such as position-based ones do not.

W2: An incorrect task-ID prediction may not imply that the action target is ambiguous.

A: Thank you for raising this important distinction. It is correct in principle, and Proposition 1 shows that this case is harmless: if the tasks compatible with a prefix prescribe the same action, reweighting leaves the learned target unchanged, and the prefix only receives a smaller gradient share. We also verified empirically that the case is rare (Sec. 5.4, App. G). Within the same task and prefix position, misidentified prefixes have higher action error, and on Walker-Param the target actions of the candidate tasks disagree 3–8× more than for correctly identified prefixes. Fewer than 6% of misidentified prefixes are assigned a task whose target action is close to the true one. Incorrect identification therefore marks ambiguous action targets in the vast majority of cases.

W3: The gains are not uniformly strong in Table 1.

A: Thank you for this observation. We respectfully disagree that it weakens the results. TIARA ranks first in five of the six environments and second in DarkRoom. It is the only method in the top two on all six, with an average rank of 1.17, versus 2.33 for DPT and 4.0 or worse for all other baselines.

DarkRoom has a structural explanation. All tasks share the same dynamics and the reward is zero until the goal is reached, so before that point the context carries no task evidence for any method. No prefix weighting can add evidence in this regime: if no prefix is identified, Eq. (11) gives uniform weights and TIARA reduces to DPT, which matches the observed difference within one standard deviation.

Q1: Have the authors considered a soft reliability score, such as entropy or confidence margin, instead of binary correctness?

A: Thank you for the suggestion. Yes, Table 2(b) compares binary correctness with three soft signals: the true-task probability, the entropy confidence, and the confidence margin. Binary correctness performs best in both environments. The soft signals recover part of the gain on Reach but stay close to uniform weighting on Walker-Param. The reason is that entropy and margin do not use the true task, so a prefix confidently assigned to a wrong task still receives a high weight, whereas the true-task probability separates identified from unidentified prefixes only gradually. The details are in App. F.

Q2: Do the results change when the context and query samples are forced to be disjoint?

A: Thank you for carefully checking the details. The overlap is rare during training and cannot occur at deployment. Over 10510^5 samples from the actual loader, the query transition falls inside the context in only 0.99% of examples on Reach and about 0.5% elsewhere, while DarkRoom uses a separate query array (App. D.2).

More importantly, exploiting such overlaps would hurt rather than help at test time. If the model learned to copy the target from the context, it would rely on a shortcut that never exists at deployment, where the context consists only of the agent’s own past transitions, whose quality is not guaranteed. A model that depends on this shortcut would therefore degrade at test time. Instead, TIARA performs well on unseen tasks with self-collected contexts (Table 1), which indicates that it does not rely on such shortcuts. Finally, all DPT-style baselines use the same loader, so the overlap does not favor TIARA.

We appreciate your comments, which have helped us improve the paper.

Reviewer hHx9: are you convinced, particularly as regards novelty?

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

TIARA augments offline in-context reinforcement learning with a causal context encoder and an auxiliary classifier for training-task identity. It uses whether that classifier is correct at each context prefix as a reliability signal, downweights action losses on incorrectly identified prefixes, and decodes actions from the context representation together with a query state. Experiments cover six multi-task environments and a high-dimensional humanoid benchmark, with ablations of the encoder and reliability weighting.

Critical strengths
  1. TIARA makes a concrete architectural distinction between history-only task information and query-conditioned action prediction. The positive lower bound on the prefix weights preserves learning from ambiguous prefixes, and the normalization keeps the total action-loss scale fixed. These are sensible design choices for sparse or delayed task evidence.

  2. The empirical results are strong across varied control settings. TIARA improves over DPT on five of six main environments and reaches 800.97 ± 172.50 on Humanoid-Dir, with useful ablations showing that both the encoder and the weighting improve Reach and Walker-Param. The authors also report the DarkRoom case where DPT is slightly better, which makes the evaluation more credible.

Critical weaknesses
  1. Classifier correctness is only an indirect and closed-set proxy for identifiability. The auxiliary head predicts one of the training-task IDs, but the test tasks are unseen and the head is discarded. A prefix can be misclassified while still containing exactly the information needed for the query action, or classified correctly using features irrelevant to the query. Conversely, the weight is computed from a hard argmax, so it does not use uncertainty or distinguish a barely correct prediction from a confident one. The paper does not report calibration against an action-ambiguity measure, posterior task uncertainty, or conditional action error. Thus the gains could reflect an auxiliary representation regularizer or a preference for later/easier prefixes rather than reliability-guided action learning.

    Moreover, the idea of using decoupled auxiliary loss for training context representations have been extensively explored in Context-Based Meta-Reinforcement Learning literature like FOCAL, UNICORN. In particular, the UNICORN papers introduced the exact cross-entropy loss and also showed similar implementations by appending the context token to a causal decision transformer. While this paper presents a more comprehensive sturdy of such mechanisms in ICRL, the technical novelty is limited.

  2. The training context order appears mismatched to deployment. Appendix D.2 states that the sampled context transitions are randomly permuted while preserving each complete transition; prefix indices therefore refer to the presented random order. At deployment, Algorithm 2 appends episodes and retains the most recent complete transitions in chronological order. This changes which evidence is available at each prefix and hence changes both the task-classification signal and the action targets that the weighting is intended to regulate. The reported ablation does not compare chronological training with shuffled training, and the paper does not show that the reliability curve survives this distribution shift.

Questions for authors
  1. Identifiability validation: Can you report task-classification accuracy/calibration, action error, and an ambiguity measure as functions of prefix length, separately for correct and incorrect prefixes? How does a soft uncertainty weight or posterior entropy compare with the hard argmax indicator in Eq. (9)?

  2. Ordering control: What are the results when contexts are trained in chronological order, with the same data and update budget, versus the random permutation in Appendix D.2? Does TIARA still outperform DPT when training and deployment use the same ordering?

  3. Mechanism ablation: Can you compare (i) the auxiliary task loss with uniform action weights, (ii) reliability weighting from a frozen classifier, and (iii) reliability weighting from an action-ambiguity or confidence signal? This would separate representation regularization from the proposed weighting mechanism.

2 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. All changes in the revised PDF are marked in blue. We number the Critical Weaknesses W1–W2 and the Questions Q1–Q3 in review order, splitting W1 into (a)–(c) by topic.

W1(a), Q1: Does binary correctness reliably reflect identifiability and action reliability, and how does it compare with soft uncertainty weights?

A: Thank you for this important question. ρ\rho is not intended to measure intrinsic identifiability. It is a training-time signal of whether the current representation identifies the task, which is exactly the information the action decoder can use. Proposition 1 (Sec. 3.3) makes its effect precise: when the tasks compatible with a prefix prescribe the same action, reweighting does not change the learned target, and binary weights do not produce a binary target, so a barely correct prediction is not treated as certain.

The measurements you request are now in Sec. 5.4 and App. G, and they confirm that ρ\rho tracks reliability:

  • Accuracy and calibration: accuracy rises from 52% at one transition to over 97% after 20, and the classifier is calibrated on short prefixes. Hence ρ\rho follows the amount of task evidence in a prefix, and a correct argmax on a short prefix reflects calibrated confidence rather than a lucky guess.
  • Action error and ambiguity: at the same position, misidentified prefixes have higher action error and 3–8× more ambiguous action targets than correctly identified ones. Hence ρ=0\rho=0 marks exactly the prefixes whose action supervision is unreliable, which is what our weighting down-weights.
  • Soft weights comparision: binary correctness outperforms the true-task probability, entropy and margin (Table 2(b)).

W1(b), Q3: Are the improvements truly attributable to reliability-guided weighting, rather than auxiliary representation learning or a preference for later prefixes?

A: Thank you for suggesting these controls. Yes, the gain comes from the weighting, as the extended ablation in Sec. 5.3 (Table 2) shows:

  • (i) Auxiliary loss. TIARA outperforms uniform weights with the same architecture and auxiliary loss ("w/o Weight") by 123.4 on Reach and 81.6 on Walker-Param.
  • (ii) Frozen classifier. TIARA outperforms a frozen classifier with the same initialization and budget by 47.8 on Reach and 93.5 on Walker-Param (App. F), so the signal must follow the representation.
  • (iii) Confidence-based signals. TIARA outperforms the true-task probability, entropy and margin by 20.5–41.2 on Reach and 73.7–92.1 on Walker-Param.
  • Preference for later prefixes. TIARA outperforms position-based weights ("w/o Encoder") by 91.7 on Reach and 47.5 on Walker-Param. By Proposition 1, position weights only redistribute capacity across prefixes, whereas correctness weights change the learned target.

W1(c): Decoupled auxiliary losses have been explored in FOCAL and UNICORN; is the technical contribution limited?

A: Thank you for the pointers to FOCAL and UNICORN. Learning a task representation from context with auxiliary supervision is a standard building block in context-based meta-RL and ICRL, which we adopt deliberately as infrastructure. Even so, our representation differs in how it is learned and used: FOCAL and UNICORN infer a single task embedding from a context to condition the policy, whereas TIARA's causal encoder produces a supervised representation for every prefix and feeds it into a DPT-style decoder through multi-layer fusion. Our main contribution is to use the per-prefix identification outcome to modulate action supervision, which targets a concrete mismatch in DPT-style offline ICRL, as Reviewer jZq8 also notes. In the ablation above, "w/o Weight" applies the UNICORN-style auxiliary loss within our architecture, and the weighting on top of it provides most of the gain; Proposition 1 shows that it changes the learned target, which an auxiliary loss alone does not.

W2, Q2: Does the mismatch between permuted training contexts and chronological deployment contexts affect reliability and generalization?

A: Thank you for pointing this out. We respectfully note that permuting the context during pretraining is the standard protocol of DPT (Lee et al., 2023), and all DPT-style baselines use the same loader, so Table 1 compares all methods under the same orderings. Moreover, intermediate chronological prefixes are never executed at deployment: the context is updated only between episodes and the action is read from the last position (App. B.2), so after the first episode (two on Reach) every decision uses the full context. Since each token is a complete transition (s,a,s′,r)(s,a,s',r), a full context carries the same task evidence regardless of order. A chronological-training control would thus mainly change the positional statistics seen by the encoder, not the comparison with DPT.

Thank you again for the helpful comments.

Reviewer bnUD: are you convinced by the rebuttal, particularly the argument about classifier correctness and the new measurements?

Comments on the paper

2 comments
Area_Chair_jqXEOpenReview ↗

From AC.

Dear Authors,

Given the deadline on 18 November, please submit your responses to the reviews to facilitate a productive discussion with the reviewers.

Thanks, AC

AuthorsOpenReview ↗

Summary of Rebuttal and Reviewer Updates

We sincerely thank the Area Chair and all reviewers for their time, constructive feedback, and careful evaluation of our work. Below is a brief summary of the rebuttal and its outcomes.

  • Additions. To address the reviewers' concerns, we added a theoretical characterization of what the weighting optimizes (Proposition 1, Sec. 3.3); reliability analyses showing that the correctness signal tracks action reliability (Sec. 5.4, App. G); architecture-matched ablations separating the weighting from the auxiliary loss and comparing it with soft signals (Sec. 5.3, Table 2); and sensitivity and capacity studies (App. F).
  • Reviewer discussion. Initial ratings were 2, 2, 2, current are 2, 2, 3. Several concerns were shared across reviewers:
    • Reviewer jZq8 asked whether correctness reflects action reliability, how it compares with soft signals, and for architecture-matched ablations and sensitivity analyses. jZq8 stated that the rebuttal addressed most of their main concerns and raised the score from 2 to 3.
    • Reviewer bnUD (no response yet) asked (1) whether correctness reflects identifiability and action reliability and how it compares with soft weights, and (2) whether the gains come from the weighting rather than the auxiliary loss or later prefixes. These are the same concerns that Reviewer jZq8 found addressed by the same analyses and ablations. We also responded to their questions on novelty and context order.
    • Reviewer hHx9 (no response yet) asked (1) whether misidentification implies action ambiguity and (2) how binary correctness compares with soft reliability scores. These match jZq8’s concerns on action disagreement and alternative reliability signals, which jZq8 found addressed. We also responded to their questions on novelty, uneven gains in Table 1, and context–query overlap.

In summary, we have responded to all concerns raised by the reviewers. The reviewer who engaged in the discussion confirmed that most of their concerns were resolved and raised the score. Although the other two reviewers have not yet responded, their main concerns are largely similar, and we are confident that our responses address their concerns as well.

We sincerely thank the Area Chair for coordinating the discussion and for the careful follow-up with each reviewer, and we greatly appreciate the reviewers’ thoughtful feedback, which has strengthened the clarity and completeness of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.