acceptodds
Under review as a conference paper at ICLR 2027

From Shallow to Deep: Pinning Semantic Intent via Causal GRPO

Abstract

Safety-aligned large language models (LLMs) remain surprisingly vulnerable to adversarial prefixes and prefilling attacks, where a short compliant context can steer an otherwise refusing model toward harmful continuation. We investigate this failure from a representation-level perspective and observe a phenomenon we term Semantic Representation Decay: harm-relevant information that is readily accessible at the end of a malicious query becomes substantially less stable as generation proceeds under adversarially induced contexts. Motivated by this observation, we propose Two-Stage Causal-GRPO (TSC-GRPO), a post-training framework for robust safety recovery. In Stage 1, we construct controlled contextual interventions that preserve the underlying harmful semantics while varying prefixes, adversarial perturbations, and partial generations. A lightweight probe is trained to learn a context-invariant harm representation, while remaining sensitive to genuine transitions from harmful continuation to refusal. In Stage 2, the frozen representation model serves as a semantic scorer during GRPO. By measuring how long each sampled continuation remains aligned with a fixed harmful-semantic anchor, we define a dense cumulative trajectory penalty that assigns higher reward to earlier safety pivots. Experiments show that TSC-GRPO significantly outperforms baselines in defending against jailbreak attacks while preserving general utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Safety Alignment

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_oCdJOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers the problem of semantic representation decay in safety training for LLM. While most existing papers apply training on offline benign and harmful prompt directly, the paper focus on the weaknesses of these approaches, which is semantic representation decay. The harmful prompt becomes unseperatable with linear probe as completion got generated. For solving this issue, the authors propose a two stages approach which in the first stage learns an invariant representation of the harmful intent, and pining on the intent during GRPO training in the second stage. Through experiment results, the authors demonstrated the approach with four different models and eight jailbreak attacks on AdvBench.

Critical strengths
  1. TSC-GRPO addresses the issue with compliant trajectory through GRPO with rollouts from compromised context directly. This directly brings robustness against these kinds of attacks.
  2. The experiments setup considers multiple model families and sizes with 8 jailbreak techniques. The setup is comprehensive and empirically enough to support some statements in the paper.
Critical weaknesses
  1. In stage 1, the learned invariant representation never learns about harmfulness, but rather than learning query identity. z_anchor is an embedding of similar queries across harmful continuation context, not really a harmful intent. This means one can still hack the reward with any drift away from the query's embedding. The robustness against unseen attacks becomes unclear.
  2. The advantage over a standard GRPO with harm-classifier reward is not clear. Since the approach doesn't learn identifying the harmfulness, training a classifier on the same dataset with harmful or benign as labels could provide even more discriminatory information.
  3. The learned invariant representation might miss-anchored a benign prompt which starts with a harmful-style prompt. Considering a multi-turn conversation starts with a harmful prompt, but gradually move back to benign. With the proposed approach, the model will not be able to change its intent representation gradually, which leads to over-refusal.
Questions for authors

1.HOw does TSC-GRPO compare to GRPO from the same forced contexts using a harmful classifier as reward? 2. Are the +- values standard deviation over multiple runs? which judge determines ASR? 3. If a model was used as the judge for ASR, a further experiments on different judges would make the result more reliable.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

We thank the reviewer for the insightful comments on the learned representation and the reliability of the trajectory reward. Stage 1 encourages representation consistency across contextual views of the same query while preserving distinctions among queries. The independent validation in Appendix C provides evidence that the learned representations retain harm-relevant information. Under a strict query-level split, the probe achieves 94.24% harmful/benign classification accuracy on unseen queries and 92.6% accuracy under compliant-prefix perturbations. These results support the preservation of harm-relevant information beyond individual training queries.

In Stage 2, the query-specific semantic anchor provides a stable reference for evaluating the evolving generation trajectory. The cumulative reward penalizes sustained above-threshold similarity to the harmful-semantic anchor, encouraging earlier recovery from harmful continuations.

To further examine reward reliability and potential semantic drift, we conduct a paired analysis on 50 held-out AdvBench queries using the frozen Stage-1 scorer. For each query, we compare three types of continuations under the same forced context and semantic anchor: harmful continuation, safe refusal, and off-topic unsafe continuation. We use independent safety annotations to verify the continuation categories and control response lengths when comparing cumulative rewards.

Table R3: Pairwise reward preference on held-out AdvBench queries.

Pairwise ComparisonRefusal Receives Higher Reward (%) ↑
Safe refusal vs. Harmful continuation86.0
Safe refusal vs. Off-topic unsafe continuation80.0

The reward favors safe refusal over harmful continuation in 86.0% of paired cases and over off-topic unsafe continuation in 80.0%. These results show that the trajectory reward generally favors safety recovery over both sustained harmful generation and unsafe semantic drift in the evaluated cases.

Together with the held-out representation validation in Appendix C and the trajectory-reward ablation in Figure 2(b), this analysis provides complementary evidence for the effectiveness of the proposed semantic anchoring mechanism.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 1

We thank the reviewer for suggesting a comparison with harmfulness classifier rewards. The key distinction lies in what the reward evaluates and how it guides optimization. A conventional response-level harmfulness classifier evaluates the safety of the final output. In contrast, TSC-GRPO focuses on safety recovery after generation has entered a high-risk state. Two trajectories may receive the same final safety label but differ in when they stop harmful continuation. Our cumulative trajectory reward captures this temporal difference by measuring the similarity between intermediate representations and the original harmful-semantic anchor. This encourages the policy to interrupt harmful continuation earlier.

This design builds on the context-invariant representation learned in Stage 1. By aligning different contextual views of the same query, Stage 1 provides a stable semantic reference for evaluating intermediate generation states. Combining semantic anchoring with cumulative trajectory feedback enables TSC-GRPO to target the recovery process underlying Shallow Safety Alignment.

The existing ablation results further support the contribution of this reward. As shown in Figure 2(b), removing the trajectory reward (α=0\alpha=0) substantially increases ASR, while keeping GRPO training and exposure to high-risk contexts unchanged. This indicates that the improvement is not solely due to GRPO or adversarial context exposure. While this result supports the effectiveness of our trajectory reward, a direct comparison with harmfulness classifier rewards under identical training conditions remains necessary to establish their relative performance. We will clarify this distinction and the scope of our empirical evidence in the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness 3

We thank the reviewer for raising this important concern about over-refusal and changes in user intent. Intent Pinning encourages semantic consistency under intent-preserving contextual perturbations. As defined in Section 4.3, h=f(c,s)h=f(c,s), where cc denotes request semantics and ss represents contextual factors. When user intent changes from harmful to benign, the semantic content cc also changes and is not subject to the same invariance constraint.

Our training design also considers benign requests. Stage 1 includes both harmful and benign queries, while Stage 2 retains standard on-policy sampling for benign queries. The semantic anchor provides trajectory-level training feedback for a given harmful request, while the trained policy responds to the current conversation context during inference.

To directly examine over-refusal, we compare Qwen2.5-7B-Instruct before and after TSC-GRPO. We evaluate the 250 safe prompts from XSTest and 30 paired multi-turn conversations in which user intent gradually shifts from harmful to benign. Each multi-turn conversation is paired with the same final benign request presented without the preceding harmful context. Both models use identical evaluation settings. The over-refusal rates are reported below.

Table R4: Over-refusal analysis on benign requests and intent transitions.

Evaluation SettingPre Over-refusal (%) ↓\downarrowPost Over-refusal (%) ↓\downarrow
XSTest Safe (250)8.88.4
Benign-only (30)6.76.7
Harmful → Benign (30)13.313.3

As shown in Table R4, the over-refusal rate slightly decreases from 8.8% to 8.4% on XSTest Safe. In the paired multi-turn evaluation, the rates remain unchanged after TSC-GRPO under both benign-only (6.7%) and harmful-to-benign (13.3%) settings. Although preceding harmful contexts increase over-refusal, TSC-GRPO does not further amplify this effect in our evaluation. These results provide preliminary evidence that TSC-GRPO preserves the ability to respond appropriately to benign requests, including those following changes in user intent.

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

This paper investigates the issue of shallow safety alignment, where LLMs generate harmful responses when the beginning of a reply is filled with affirmative phrases such as "Sure, here is".

The authors train a linear probe on the final hidden layer of the LLM to diagnose this problem and propose a two-stage training framework as a solution based on observed phenomena. Specifically, they first train a lightweight probe to align representations of different variants of the same query, then design a GRPO-based training strategy to train a policy for defense. Extensive experiments demonstrate the defense effectiveness of the proposed method.

Critical strengths
  1. This paper is well-motivated and logical. The authors first perform representational-level diagnostics and then train a probe to provide reinforcement-learning signals.

  2. The reward design is interesting. The cumulative reward mechanism distinguishes the model's trajectory of revisiting at different steps, encouraging the model to return earlier to avoid outputting unsafe content.

  3. The experimental results demonstrate that while ensuring enhanced security, the performance in areas such as reasoning and coding remains largely unaffected.

Critical weaknesses
  1. Based on my understanding, R_causal accumulates as the token length increases, inherently imposing stricter penalties on longer responses. This may induce the model to output shorter replies or lead to premature termination even for benign requests. The paper does not analyze this problem with respect to response length.

  2. The experimental model is outdated. More recent models after Qwen2.5 were not considered, which limits the algorithm's generalizability.

  3. The authors are encouraged to supplement a pipeline diagram to better explain the algorithm and enhance expressiveness.

  4. The baselines considered in the paper are also outdated, as numerous papers on safety alignment have been published since then. The authors should incorporate comparisons with more safety alignment strategies, such as [1,2,3]. Besides, [4] is a highly relevant work that can be discussed in terms of its similarities and differences with the issues addressed in this paper.

[1] Efficient Safety Alignment of Language Models via Latent Personality Traits. In COLM 2026.

[2] AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning. In ICLR 2026.

[3] TriPlay-RL: Tri-Role Self-Play Reinforcement Learning for LLM Safety Alignment. In ACL 2026.

[4] Revisiting backdoor attacks on LLMs: A stealthy and practical poisoning framework via harmless inputs. In EMNLP 2026.

Questions for authors

See weaknesses.

10 comments
AuthorsOpenReview ↗

Response to Weakness 1

We thank the reviewer for the thoughtful comment on response length and premature termination. The cumulative reward penalizes the duration of above-threshold alignment with the harmful-semantic anchor. At each generation step, a penalty is incurred only when the similarity exceeds the threshold τ\tau. Otherwise, the penalty is zero. Continued generation incurs no additional causal penalty as long as the similarity remains below τ\tau. This design encourages earlier recovery from harmful trajectories without uniformly penalizing response length.

The training objective also balances safety recovery with response quality. In Stage 2, high-risk contexts are constructed for harmful queries, while benign queries retain standard on-policy sampling. The trajectory reward is combined with a general reward to encourage response quality.

To examine the potential impact on response length, we evaluate Qwen2.5-7B-Instruct before and after TSC-GRPO on 100 held-out benign queries from Alpaca, using identical decoding settings.

Table R1: Response length and completion analysis on benign requests.

ModelMean LengthMedian LengthP10 LengthPremature Termination (%) ↓Task Completion (%) ↑
Qwen2.5-7B-Instruct (Pre)164124392.092.0
Qwen2.5-7B-Instruct + TSC-GRPO (Post)161121382.091.0

Evaluation: 100 held-out benign queries from Alpaca. Both models use identical prompts, decoding settings, and maximum generation lengths.

As shown in Table R1, the mean response length decreases slightly from 164 to 161 tokens (1.83%), while the median and 10th-percentile lengths remain comparable. The premature termination rate remains unchanged at 2.0%, and task completion decreases marginally from 92.0% to 91.0%. These results show no substantial response shortening or increase in premature termination on the evaluated benign requests, consistent with the largely preserved general capabilities reported in Table 3.

AuthorsOpenReview ↗

Response to Weaknesses 2 & 4

We thank the reviewer for these valuable suggestions. Regarding model generalizability, our existing experiments cover four models from two families: LLaMA-2-7B-Chat, LLaMA-3.1-8B-Instruct, Qwen2.5-7B-Instruct, and Qwen2.5-14B-Instruct. Table 1 shows consistent robustness improvements against eight jailbreak attacks, while Table 3 demonstrates largely preserved general utility. These results support the effectiveness of TSC-GRPO across different model families, generations, and parameter scales.

To further evaluate generalizability to newer-generation models, we extend TSC-GRPO to Qwen3-8B using the same two-stage training framework. We compare the original and TSC-GRPO-trained models under identical evaluation settings, with non-thinking mode enabled for both. We evaluate robustness against prefilling and ReNeLLM attacks on AdvBench and general utility on GSM8K.

Table R2: Generalization to Qwen3-8B.

ModelPrefilling ASR (%) ↓ReNeLLM ASR (%) ↓GSM8K (%) ↑
Qwen3-8B (Pre)49.539.886.8
Qwen3-8B + TSC-GRPO (Post)17.615.285.1

As shown in Table R2, TSC-GRPO reduces ASR by 31.9 and 24.6 percentage points against prefilling and ReNeLLM attacks, respectively, while GSM8K accuracy decreases by 1.7 percentage points. These results extend the evidence for TSC-GRPO to Qwen3-8B, demonstrating improved robustness with largely preserved mathematical reasoning performance.

We also appreciate the reviewer's suggestions regarding recent safety alignment methods. Latent Personality Alignment (COLM 2026) improves safety through adversarial training of personality-related representations, AlphaAlign (ICLR 2026) employs verifiable safety and helpfulness rewards, and TriPlay-RL (ACL 2026) introduces collaborative optimization among attackers, defenders, and evaluators. TSC-GRPO addresses safety recovery under adversarial generation contexts through content-preserving contextual interventions and cumulative trajectory-level rewards. This design encourages the policy to recover earlier from harmful continuations. We will incorporate these studies and clarify their methodological differences in the revised manuscript.

The recent EMNLP 2026 work on backdoor attacks via harmless inputs is also relevant, as it examines how harmful behavior can be sustained during generation. While that work aims to maintain harmful compliance, TSC-GRPO seeks to promote recovery from compromised generation contexts. We will discuss this connection in the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness 3

We thank the reviewer for the valuable suggestion. We will add an overview figure in the revised manuscript to illustrate the two-stage TSC-GRPO framework and its key components.

The figure will present the contextual view construction, frozen encoder, and probe optimization in Stage 1, followed by high-risk context sampling, semantic anchoring, cumulative trajectory reward computation, and GRPO optimization in Stage 2. It will also highlight how the context-invariant representations learned in Stage 1 provide stable semantic signals for trajectory-level optimization in Stage 2.

This addition will clarify the overall workflow and the connection between the two stages, making the proposed method easier to understand.

Reviewer_8ht8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies recovery after an adversarial context has already steered generation toward harmful compliance. It observes that a fixed query-level harmfulness probe becomes less accurate as generation proceeds. TSC-GRPO first trains a representation probe on contextual variants of the same request. It then freezes the probe and backbone to score continuations during policy training. Similarity to the original harmful query is accumulated into a penalty. GRPO uses this penalty together with a general quality reward. Experiments cover four models, jailbreak attacks, fine-tuning attacks, and general utility.

Critical strengths
  • Recovery is a valuable training objective. Teaching models to stop after harmful generation has already begun is a valuable goal. Training from forced prefixes and partial harmful responses directly addresses this setting. The cumulative penalty is intended to favor earlier recovery. This is a concrete and useful objective beyond an initial refusal.

  • The experiments provide encouraging evidence for the method. Results across four backbones show improved robustness with largely preserved general utility. The control with the causal reward removed keeps adversarial exposure and GRPO training in place. Its worse performance supports the contribution of the reward beyond exposure alone. Held-out probe validation also avoids sharing views of the same query across training and validation.

Critical weaknesses
  • W1. What teaches the scorer to recognize recovery? As I understand the method, the reward assumes that moving away from the harmful query's representation (the anchor) means recovering safety. Yet a refusal such as "I UNDERSTAND WHAT YOU ARE ASKING, but I cannot help with that" may still preserve information about the original request. Recognizing harmful intent and complying with it are different properties. With normalized vectors, the trajectory reward is

    Rcausal(y)=−∑t=1∣y∣max⁡ ⁣(zt⊤zanchor−τ,0).R_{\mathrm{causal}}(y)=-\sum_{t=1}^{|y|}\max\!\left(z_t^\top z_{\mathrm{anchor}}-\tau,0\right).

    Here ztz_t represents the context after token tt, and τ\tau is a similarity threshold. The introduction mentions a recovery-separation objective, but the stated losses only align contextual views and discourage representation collapse. The training description does not include refusing continuations of harmful queries. Current probe validation tests harmful/benign classification, including under compliant prefixes. The qualitative examples also illustrate late refusal. Neither establishes that refusal reduces similarity to its own harmful anchor. This leaves the interpretation of the central reward insufficiently supported.

  • W2. The proof of Theorem 1 has two gaps. The proof shows that zero alignment loss is achievable, but not that it is optimal when entropy is also optimized. It also assumes that high entropy preserves all semantic distinctions. Having many different output values does not guarantee a distinct value for each distinct meaning. Please justify these steps or weaken the theorem.

Questions for authors

Q1. Was a "recovery-separation" loss used? If so, please provide the loss and training examples. Otherwise, how do the stated losses teach the scorer to distinguish harmful continuation from refusal? Please compare both continuations for the same queries. Plot similarity to each query's anchor over generation alongside independent (manual) safety judgments. This would directly test the proposed reward interpretation.

Q2. Can you justify the two steps identified in the proof of Theorem 1, or narrow the identification guarantee? Please distinguish what the theorem establishes from what the practical training objective encourages. (See W2)

W3. Table 1 identifies new measurements and results taken from other papers, but their evaluation protocols remain unclear. Please specify the harmfulness judge, attack budgets, and stopping rules. Were attacks regenerated against each defended model? Please identify which comparisons use the same protocol and qualify superiority claims where conditions differ.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1 & Question 1

We sincerely thank the reviewer for this insightful comment. We agree that recognizing the original harmful intent and continuing to comply with it are distinct properties. This distinction is central to our Intent Pinning design. Our goal is to preserve harm-relevant semantic information under adversarial contextual perturbations while allowing the representation to reflect changes in generation behavior. A safe refusal does not require forgetting the original harmful request. It represents a change in how the model responds to that request.

Stage 1 uses the alignment and KoLeo uniformity losses in Eqs. (5)–(7), without a separate recovery-separation loss. We will clarify the corresponding wording in the Introduction. The alignment objective is applied to intent-preserving contextual views, including compliant prefixes, adversarial suffixes, and partial harmful continuations. As explained in Appendix G.2, refusal transitions are not included as positive pairs for alignment. Therefore, Intent Pinning preserves the semantic reference across harmful-intent-preserving contexts without imposing the same invariance constraint on safety recovery.

In Stage 2, the semantic anchor is computed from the original harmful query, while each intermediate representation is computed from the complete context, including the generated continuation. The trajectory reward is defined as

Rcausal(y)=−∑t=1∣y∣[cos⁡(zt,zanchor)−τ]+.R_{\mathrm{causal}}(y)=-\sum_{t=1}^{|y|}\left[\cos(z_t,z_{\mathrm{anchor}})-\tau\right]_+.

This reward measures the evolving generation trajectory's residual alignment with the harmful-semantic anchor, rather than simply classifying whether the original request is harmful. Thus, a refusal may retain the original request information while changing the representation of the ongoing response. The cumulative penalty encourages the policy to reduce prolonged alignment with harmful continuations and recover toward refusal earlier.

Our existing experiments support the effectiveness of this reward mechanism. Appendix C reports 94.24% held-out validation accuracy and 92.6% accuracy under compliant-prefix perturbations, supporting the stability of the learned harm-relevant representation. More importantly, Figure 2(b) shows substantially higher attack success rates when the trajectory reward is removed (α=0\alpha=0), despite retaining the same GRPO training setting and adversarial contexts. Figure 3 further demonstrates that TSC-GRPO can recover toward refusal under forced compliant prefixes where the original model continues harmful generation. Consistently, Table 1 reports 0.00% prefilling attack success rates across all four evaluated models. These results support the contribution of the learned semantic representation and trajectory-level reward to safety recovery.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 2

We sincerely thank the reviewer for carefully examining Theorem 1 and identifying these two points. We agree that the proof should distinguish the invariance established by the alignment objective from the additional role of entropy regularization. We will refine the theorem and its proof accordingly.

Regarding the first point, achieving zero alignment loss does not necessarily imply a global optimum of the joint alignment–entropy objective, since the entropy term may introduce a trade-off. We therefore narrow the theoretical conclusion to the contextual invariance established under exact alignment. Specifically, when

Lalign(g)=E[∥g(f(c,s))−g(f(c,s~))∥22]=0,\mathcal{L}_{\mathrm{align}}(g)=\mathbb{E}\left[\|g(f(c,s))-g(f(c,\tilde{s}))\|_2^2\right]=0,

the non-negativity of the squared distance implies

g(f(c,s))=g(f(c,s~))g(f(c,s))=g(f(c,\tilde{s}))

almost everywhere over the intervention pairs. Under sufficient intervention coverage, this equality makes the representation invariant across the contextual states connected by the interventions. On a connected contextual support for each fixed semantic content, it follows that

g(f(c,s))=ψ(c)g(f(c,s))=\psi(c)

almost everywhere. This establishes the contextual invariance property without assuming that the joint objective necessarily attains zero alignment loss.

Regarding the second point, we agree that maximizing entropy does not by itself guarantee an invertible mapping of semantic content. The entropy term discourages representational collapse and promotes diversity among learned representations, but does not establish one-to-one identification of distinct semantic contents. Accordingly, we will remove the invertibility requirement on ψ\psi and state the theorem as a conditional contextual invariance result rather than a guarantee of complete semantic identification.

This distinction is also consistent with our practical implementation. Stage 1 uses the alignment loss over content-preserving contextual views together with the KoLeo uniformity regularizer. The former encourages invariance to contextual perturbations, while the latter discourages collapsed representations. The theoretical result characterizes the invariance property under idealized intervention coverage and exact alignment, whereas the practical objective encourages this property through finite-sample optimization. TSC-GRPO relies on the resulting stable harm-relevant representation for trajectory-level scoring, without requiring exact recovery of the latent semantic variable. The training procedure and reward formulation remain unchanged.

AuthorsOpenReview ↗

Response to Weakness 3

We sincerely thank the reviewer for this valuable comment. We clarify the evaluation settings used in Table 1 as follows.

For harmfulness evaluation, we employ the HarmBench Classifier(HarmBench-Llama-2-13b-cls) to assess attack success based on whether the generated response fulfills the harmful request. The same classifier and evaluation criterion are consistently applied across all evaluated models and attack methods in our experiments.

For attack budgets, GCG uses a 20-token adversarial suffix, top-k=256, 512 candidates per iteration, and a maximum of 500 iterations. AutoDAN uses 100 optimization steps with a batch size of 64. PAIR uses 5 attack streams and up to 5 refinement iterations, with at most 5 attempts per iteration. ReNeLLM uses up to 20 evolution steps, while ICA uses 5 in-context demonstrations. CodeChameleon, DeepInception, and prefilling attacks use their predefined encoding, nested-context, and forced-prefix constructions, respectively. Iterative attacks terminate upon satisfying their success criteria or reaching the maximum budget, while fixed-construction attacks do not involve iterative optimization.

For attack generation, GCG, AutoDAN, PAIR, and ReNeLLM generate or optimize adversarial prompts against each defended target model. CodeChameleon, DeepInception, ICA, and prefilling attacks use the corresponding fixed attack constructions across models.

All newly measured ASR results in Table 1 are evaluated using the same HarmBench Classifier under a consistent evaluation protocol. Previously published results are clearly distinguished from our measurements, and direct performance comparisons are restricted to matched evaluation settings. We will clarify these details in Table 1 and Appendix A.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.