acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Personalization Under Preference Epistemic Uncertainity

Abstract

Human preferences change over time: an established preference may remain valid over time, become uncertain as context shifts, or be explicitly revised. Effective personalization requires conversational AI to adapt its response strategy to evolving preferences rather than relying on passive memory retrieval. We formalize this capability as adaptive personalization under epistemic uncertainty. A human preference study further confirms that people favor state-contingent response strategies as preference uncertainty changes. To evaluate whether language models mirror this adaptation, we create a multilingual benchmark spanning 250 personas, 5 languages, and 25K conversational turns that decouples strategy selection from response execution. Across eight closed- and open-weight models, we uncover a recognition–generation gap: models reliably identify appropriate strategies when selecting among candidate options, but collapse during open-ended generation, suffering a 35-46% drop in strategy calibration. Crucially, this failure is systematic: models default to exploit-first behavior, over-relying on existing preferences regardless of preference epistemic uncertainty. We show this failure is addressable: while prompting interventions fail, targeted training that supervises models' internal strategy representations toward state-conditional human preferences narrows the user-acceptability generalization gap by a third on out-of-distribution personas and conversational styles. Code and datasets will be released upon acceptance.

Primary area:
datasets and benchmarks
Keywords:
adaptive personalization, strategy selection in conversations

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

This paper studies adaptive personalization under preference epistemic uncertainty, where LLMs must adjust their response strategies as user preferences become uncertain or change. The authors introduce a multilingual benchmark with 250 personas and approximately 25K conversational turns, revealing a substantial gap between recognizing appropriate strategies and generating responses that follow them. To address this, they propose a GRPO reward that aligns strategy distributions decoded from hidden representations with human preferences, demonstrating improvements in strategy calibration and generalization.

Critical strengths
  • The paper goes beyond preference retrieval to study whether models adapt their response strategies as preference reliability changes. The recognition–generation comparison and oracle-strategy experiments provide compelling evidence that the main difficulty lies in selecting appropriate strategies rather than executing them.

  • The benchmark covers diverse personas and languages, with human validation supporting its quality. The proposed Link reward also consistently improves strategy calibration and user acceptability across multiple models.

Critical weaknesses
  • Unclear contribution of representation-level supervision. Unlike the binary strategy calibration reward, the Link reward provides a continuous distributional training signal. It is therefore unclear whether the improvements arise from supervising internal representations or simply from providing richer strategy-level feedback. A baseline using the same KL reward computed from an output-level strategy classifier is needed to establish the specific benefit of representation-level supervision.

  • Limited causal evidence and potentially restrictive evaluation criteria. The strategy probe uses hidden activations from the first eight generated tokens, which may primarily encode strategy cues already expressed in the output rather than an internal strategy-selection mechanism. Moreover, the state-dependent binary acceptability criteria may incorrectly penalize valid responses, particularly when a model directly acts on newly revised preferences. These issues raise questions about the interpretation of both the observed failures and the proposed improvements.

Questions for authors
  • How does the Link reward compare against an output-level strategy-distribution reward using the same KL objective? This comparison would help isolate the contribution of representation-level supervision.

  • Can the strategy be reliably decoded from hidden activations before generation begins? Does Link training increase the causal influence of the identified state direction on subsequent strategy selection?

  • How are responses that directly act on newly established preferences handled in the Revised state? Do the reported findings hold under human evaluation of generated responses rather than LLM-based strategy judgments?

1 comment

We thank the reviewer for the comments and address the main questions and weaknesses below.

Question 1: How does link reward compare against an output-level strategy-distribution reward

To isolate what activations add, we trained GRPO on Gemma-3-12B with the same KL objective, human target Pref(s), λ and training budget, but with the strategy distribution predicted from the generated response by an output-level classifier instead of from activations by the probe (Link′). In-distribution, Link′ is close to Link (SC 75 vs. 77; R 3.94 vs. 3.95; UA 63 vs. 65; GRPO: 70, 3.89, 60). On the benchmark’s unseen personas, Link′ falls further behind (SC 64 vs. 70; R 3.36 vs. 3.45; UA 57 vs. 60; GRPO: 59, 3.31, 53). Its generalization gap (in-distribution minus benchmark SC) is 11, the same as GRPO’s, while Link’s is 7 (Link vs. Link′ on benchmark SC: [McNemar p < 0.05]). We conclude that an output-level distributional reward accounts for much of the in-distribution improvement, while computing the reward from activations improves generalization to unseen personas.

Question 2: Can the strategy be reliably decoded from hidden activations before generation begins?

We refit the probe on the last prompt token, right before the first reply token, at the same layer, on all four open-weight models. Strategy is decodable at 49-54% (chance 20%), versus 70.0% from generated tokens, so a signal exists before any output. We further experiment with causal ablation to address the reviewer’s point on the causal influence of the identified state direction on subsequent strategy selection. On Gemma-3-12B, over the full benchmark (5.6K prompts), removing the state subspace lowers the SC metric from 59 to 55 in the GRPO model and from 70 to 62 in the Link reward model, compared with 57 and 67 when a random subspace of the same rank is removed. The larger state-specific drop for Link (8 vs. 4 points; difference p < 0.05, paired test) is consistent with Link training increasing the subspace’s influence on strategy selection. This experiment shows supporting evidence that Link reward training does increase the model’s reliance on the preference epistemic state subspace.

Question 3: How are responses that directly act on newly established preferences handled in the Revised state?

In the Revised epistemic state, a response that acts directly on the newly stated preference is labeled recalibrate, because it updates behavior once new evidence contradicts the earlier preference. Thus, recalibrate is an acceptable strategy for this state (Table 1). Strategy exploit in Revised means continuing to act on the superseded (old) preference, which is not an acceptable strategy. To check that the findings do not depend on LLM judgments, we had two annotators label 300 generated responses from Gemma-3-12B (100 each, from the untrained, GRPO and Link reward models, on the same prompts), blind to the jury labels and to which model wrote each response. They agreed with the jury at κ = 0.74 for strategy and κ = 0.80 for acceptability. Under their labels, 38% of the untrained model’s responses, 57% of GRPO’s and 66% of Link’s were acceptable, the same ordering as our jury-based results.

We hope these responses and the additional results address your concerns. If they do, we would appreciate it if you would consider updating your assessment of the paper. We are happy to answer any further questions during the discussion period.

Reviewer_qcjLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how an assistant respond when a preference is updated . The authors find that models often choose suitable responses but keep acting on existing preferences in their own answers when clarification or adjustment would be more appropriate. They also found out that models are more likely to select right answer rather than coming up with right answer by themselves

Critical strengths

I like the question this paper asks. An assistant should be able to reconsider a preference when the current situation asks for it. Looking at whether it should act, ask, or adjust is useful. I also appreciate the experiment and the comparisons with standard training methods. The training improvements are a strength, even though I have questions about what the benchmark is measuring. I like paper carried out quite a few experiments to measure different performances across methods.

Critical weaknesses
  1. The main finding feels close to the familiar difference between recognizing a good answer and producing one. PrefEval [1] already evaluates both on the same preference queries and reports better performance when models choose from options. I see why studying this across preference states is useful, but I need a clearer explanation of what we learn beyond those findings. Since the options contain complete responses, the gap by itself does not establish why the model fails when writing its own answer.

  2. I also find it hard to separate over-reliance on an old preference from the preference-update failures already studied in PersonaMem. That benchmark asks models to follow current preferences and documents errors where they choose outdated ones. What does the uncertainty framing reveal beyond this? A new set of generated conversations is useful, but preference updates themselves are already covered there.

  3. The distinction between a changed preference and a preference that depends on the situation is not clear enough to me. If I usually book budget hotels but want a nicer one for a reunion, my general preference may not have changed. Table 7 seems close to this: it describes a three-star to four-star update around a reunion. The excerpt may leave out relevant context, but I would like to know how these cases are labeled and whether the benchmark checks that the assistant returns to the usual preference afterward.

  4. People also differ in how strictly they follow a preference and whether they want the assistant to ask again. I appreciate that the paper uses different personas and persona-specific ranking. Still, Table 1 and the training reward use a shared distribution of preferred strategies for each state. How do you check that this matches the particular person? An assistant could know with certainty that someone makes exceptions, so flexibility should not itself be treated as uncertainty. I would like to see how the labels account for this difference.

  5. I would like stronger human evaluation of the training results. The paper does include a human preference and annotation study, which I appreciate. My concern is that the training conversations and benchmark are both generated with the same pipeline, even though they use different models and personas, and the reported user-acceptability scores come from a simulator. Do people actually prefer the trained model's answers? A comparison judged by people, ideally including conversations with real users, would help show that the improvement carries beyond the synthetic setup and its labeling rules.

  6. I would also like to know whether the main findings hold for newer models. The evaluation includes GPT-5.4 and Opus 4.7, while GPT-6 Astra and Opus 5.5 are now available. Their reported progress on extended tasks makes this a useful check before drawing broad conclusions about current assistants. I am not assuming that better long-task performance fixes personalization, but the paper needs direct evidence about whether the same failures persist, with the reasoning settings made clear.

  7. The practical setting also needs more attention. Users often interact with an assistant such as ChatGPT or Claude, where memory and context management can affect the answer. For example, Anthropic describes how Claude Code uses context compaction, persistent notes, and retrieval. So, I find it difficult to tell how much of the reported problem users would encounter in a deployed assistant.

I find the study interesting, including the training results, but these issues currently make it difficult for me to judge what the benchmark adds and what its scores mean. My recommendation is borderline reject. I am open to changing it if the authors clarify these points and provide convincing support for the distinctions they make.

Reference

[1] Zhao et al. Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs. ICLR 2025.

[2] Jiang et al. Know Me, Respond to Me: Benchmarking LLMs for Dynamic User Profiling and Personalized Responses at Scale. COLM 2025.

Questions for authors
  1. What does this benchmark show that earlier work on choice versus generation and preference updates could not show?

  2. How do you distinguish a lasting change from an exception for one occasion, and what happens after that occasion ends?

  3. How do the strategy labels account for people who differ in how strictly they follow their preferences or want clarification?

  4. Do people prefer the trained model's responses, including in conversations outside the synthetic setup?

  5. Do the same failures persist on newer models such as GPT-6 Astra and Opus 5.5?

  6. What happens when the assistant can manage its own memory and search earlier conversations?

3 comments

We thank the reviewer for the comments and address the main questions and weaknesses below.

Weakness 1: I need a clearer explanation of what we learn beyond those findings

Our main findings include three new results: First, the recognition-generation gap depends on how reliable the preference is: it is +7 points when Established, but +45 when Uncertain and +36 when Revised (Section 3.2), which a setting with only valid preferences (PersonaMem) cannot show. Second, the failure lies in choosing the strategy, not in writing the response: when told which strategy to use, models comply 91–98% of the time and calibration rises from 83.1 / 36.9 / 47.9% to 96.4 / 93.1 / 94.0% (Tables 23–24), while unaided they default to exploit (Section 3.3). Third, the failure is not a lack of information: preference histories, instructions, chain-of-thought and few-shot examples do not close the gap (Section 4.1). Section 4.2 then examines the model’s internal representations and shows that a reward targeting the state–strategy link improves calibration on unseen personas. We will also add PrefEval to Related Work and Table 16.

Weakness 2: What does the uncertainty framing reveal beyond this?

PersonaMem tests whether a model uses the user’s current preference: it picks, from four options, the one consistent with the current profile, and the distractors rely on outdated or irrelevant information. Instead, our benchmark tests what the assistant should do when it is unsure how much to trust the preference: act on it, ask about it, confirm it, hedge, or update. This differs from PersonaMem in three crucial ways: (1) Uncertain evidence, not only updates. Preferences are often hinted at before they are stated, and in our Uncertain state the user has only hinted. An assistant that treats a hint as settled overcommits, and one that ignores it misses the user; models reach only 36.9% strategy calibration here, and related work like PersonaMem has no counterpart. (2) Scoring the strategy, not only the value. Humans in our preference study chose to act directly on the preference 90.4% of the time when it was Established, but only 8.5% when Uncertain and 6.2% when Revised, preferring to elicit, confirm, hedge or recalibrate (Table 1). Using the latest value is therefore not sufficient: the right behavior depends on how reliable the evidence is, whereas PersonaMem has a single correct response. (3) Free-form generation, not selection. Deployed assistants write responses and do not choose among options, and selection overstates their ability. In Revised, models pick an acceptable response 83.5% of the time in multiple choice but generate one only 47.9% of the time, defaulting to exploit in 52% of responses (Sections 3.2–3.3), so the failure is not in tracking the update.

Weakness 3: distinction between a changed preference and a preference that depends on the situation

In our benchmark, a revised preference is created as a lasting change. Each drift event is generated at plan time with an explicit {old, new} value that takes effect during a given session and overwrites the persona’s preference in all later sessions. The plan also includes at least three post-drift sessions in the affected domain, so the new preference is tested again afterward (Appendix A.3). Once the user states the new value, the state returns to established, so “revised” covers only the window in which acting on the old value is wrong. In the example in Table 7, the reunion is where the persona first mentions the change from 3-star to 4-star hotels. This is a lasting change for the persona, and the 4-star preference continues in later sessions.

In this work, we do not model temporary, situational preferences, so the benchmark does not check whether the assistant returns to the usual preference afterward. Within our framework, such a case could be expressed as two drift events: a change to the temporary value, followed by a change back to the original. We note that this is an interesting future work. We will clarify this distinction of preference shifts in Section 2 and discuss situational changes as a limitation in this work.

Weakness 4: People also differ in how strictly they follow a preference and whether they want the assistant to ask again

We address the two parts of this concern separately. First, flexibility is a persona attribute, not uncertainty. Epistemic state is defined by what the user has stated about a specific preference, computed from the jury’s revelation level and whether the stated value is current (Appendix A.3). Traits such as flexibility or tolerance for questions are separate persona attributes (friction traits and interaction-style attributes, Appendix A.2), so a flexible user is labeled uncertain only if they actually leave a preference unsettled in the conversation. Second, Table 1 and the training reward use a population-level distribution of preferred strategies per state. This is deliberate: we first align the model to what users prefer on average, and target a distribution and not a single strategy because several strategies are acceptable in each state. The model is conditioned on the persona and conversation history (Section 2.4), so it can select among them to suit the individual user. We do not analyze individual differences separately from epistemic state, and the acceptable sets pool across persona traits; we will state this as an extension/future work in the revised paper.

Weakness 5: human evaluation of the training results

Three annotators each chose their preferred response among the untrained model (Gemma-12B), GRPO and GRPO + Link on 100 English prompts [from the benchmark’s unseen personas] (300 judgments; Fleiss’ κ = 0.40). They chose Link in 51% of judgments, GRPO in 37% and the untrained model in 12%. The trained models were strongly preferred over the untrained model (p < 0.001). We will add this validation study to the appendix.

Weakness 7: memory and context management

Our oracle-memory intervention (Section 4.1) provides evidence that the model failures may not stem from limited memory or access. This intervention gives the model a structured summary of the user’s preference history (a stronger memory than a deployed system would typically have); this additional context changes strategy calibration by only −0.6 / +0.4 / +0.8 points (Established / Uncertain / Revised; Table 31). One reading is that the difficulty lies in deciding how to act on the evidence, not in retrieving it, and a memory that records the preference does not itself indicate how reliable it currently is.

We hope these responses and the additional results address your concerns. If they do, we would appreciate it if you would consider updating your assessment of the paper. We are happy to answer any further questions during the discussion period.

Thanks for the rebuttal and the additional human evaluation. I understand the contribution better now. The paper shows that the gap between picking a good response and writing one gets larger when a preference is uncertain or revised. I see what that comparison adds, but the novelty still feels limited to me.

My concern about how this works in practice also remains. Giving the model the correct preference label tests whether that information helps. I still don’t know whether the same failures would show up in an agent that manages its own memory or in a chat product with its normal memory and context handling.

I appreciate the clarifications and the new experiments. I’m keeping my score at 2 (weak reject).

Reviewer_fFcyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper defines three epistemic states: established, uncertain, and revised for each evidence regarding user preferences, where each state demands a distinct conversational strategy. The paper introduces a multilingual benchmark of 250 personas, 5 languages, and 25K turns. They find that models can identify the correct strategy, but struggle to generate a response. Models default to exploit-first behavior and over-rely on existing preferences. The authors show that training the model to align its internal strategy representation to state-conditional human preferences improves its responses.

Critical strengths

Examining the uncertainty over each preference item, and deciding what action to take is a novel and underexplored area. The benchmark can be valuable for future work.

The benchmark is multilingual, natively generated (although with prompts in English), which is a good effort to evaluate and develop LLMs outside of English.

Critical weaknesses

It is unclear whether the strategy is the right level of abstraction for deciding what to do in personalized responses under different epistemic states. Also, according to the description in the paper (L1110116), some strategies do not correspond to explicit observable actions by the user (confirm and recalibrate). So the strategies do not seem like parallel actions but rather sequential steps the model can take. For example, it can elicit then recalibrate. It would be more helpful if the authors can provide more details on how a model actually uses these strategies, and how responses are generated and validated based on these strategies.

The benchmark construction pipeline and human-in-the-loop development is not clearly documented or justified. For example, it’s unclear how human annotators code the free-response failures into the failure modes, what is the agreement. It’s also unclear how the rubric was developed. The construction of the options (L158-161) is not well-justified. The authors include two acceptable strategies and two negative options conditioned on failure modes that do not necessarily conflict with the acceptable strategies, and there is no unacceptable strategy in the options. More questions on clarity below.

In the training section, the probe is trained to predict the strategy from the initial tokens of the response, so adding the KL against this probe to the reward is essentially teaching the model to approximate the human preference distribution, which is leaking more information to the model than standard GRPO. A more reasonable baseline that isolates the “aligning internal state” claim should use a LLM-judge to predict strategy, then use the same KL to Pref(s).

Questions for authors

In the human study of 500 dialogue instances, how do the authors validate that the LLM outputs actually embody the strategy that it’s linked to?

“in each epistemic state, a small set of strategies account for the vast majority of preferences (> 80%) (Table 1), suggesting a natural separation between acceptable and unacceptable strategies.” (L132-135)

This is not really true, according to Table 1, in Uncertain and Revised, 3 out of 5 and 4 out of 5 strategies account for the majority of preferences. Even then, how does this suggest a natural separation between acceptable and unacceptable strategies?

“In every state, one split fit the observed preferences best.” (L139)

This is just saying the splits are strictly ranked, but does not provide information on whether the best split is statistically significantly better than the second ranked split.

Does the motivational human study and the validation use the same 500 dialogues? What is the human agreement of the human study in S2.2?

It is unclear what models are used in the LLM jury and rubric scorer in S2.4.

Rubric development is not well-justified, the rubric dimensions described by the paper directly conflict with some strategies. For example, rubric has confidence, which a hedging or elicit strategy should not embody. Since Established has Exploit as the main strategy, it then makes sense that it has higher correlation with rubric judgments in S3.4. Further the strategies come from user ratings, so it’s circular to correlate user acceptability to strategy calibration in S3.4.

2 comments

We thank the reviewer for the comments and address the main questions and weaknesses below.

Weakness 1: Whether the strategy is the right level of abstraction

Our unit is the assistant’s move in a single turn, defined relative to its belief about the preference. The strategies are not mutually exclusive in practice: an assistant can elicit and later recalibrate, or combine moves in one response. The user-side variable is the epistemic state, and the strategy is the assistant’s response to it. Because several strategies are acceptable in each state (Table 1), we score acceptability against a set and not a single strategy. Benchmark responses are generated to implement one strategy and validated by human annotators, who label which strategy each response uses (κ = 0.71, Table 13). When models are told which strategy to use, they comply 91–98% of the time (Table 24), so each strategy can be executed as a distinct action. We have added a paragraph on combined and sequential strategies to Section 2, with definitions, examples and the generation and validation pipeline in Appendix A.1 and Table 3.

Weakness 2: Benchmark construction pipeline and rubric development

The rubric measures execution quality independently of strategy choice, so that selection (SC) and execution (R) can be separated. Its five binary criteria (Information, Confidence, Intent, Scope, and Style) ask whether a response provides the relevant information, conveys a level of certainty appropriate to the epistemic state, addresses the user’s intent, stays within the scope of the request, and matches the user’s stated interaction style. We designed these criteria to capture whether a response is both correct and appropriately personalized, and we do not claim they are the only valid decomposition. We will add the full rubric prompts in the Appendix. Moreover, the rubric does not favor exploit or Established: when the strategy is fixed by instruction, R is similar across states (3.58, 3.52 and 3.56) and across strategies (3.44–3.61; Tables 23–24), so the stronger SC–R correlation in Established is not explained by the rubric rewarding exploit. The within-state correlations are computed over eight models and are descriptive (Appendix B.6).

Weakness 3: Training baseline using an LLM-judge to predict strategy, then the same KL to Pref(s)

We trained an additional baseline as reviewer describes (Link′): the same KL objective, Pref(s), λ and training budget, with the strategy distribution predicted from the response by the open-weight LLM jury used during training (Appendix B.8) instead of using the activation probe. On Gemma-3-12B, Link′ recovers much of the in-distribution gain (SC 75 vs. 77 for Link and 70 for GRPO; R 3.94 vs. 3.95 vs. 3.89; UA 63 vs. 65 vs. 60). On the benchmark’s unseen personas it falls further behind (SC 64 vs. 70 vs. 59; R 3.36 vs. 3.45 vs. 3.31; UA 57 vs. 60 vs. 53), and its generalization gap (in-distribution minus benchmark SC) matches GRPO’s (11), while Link’s is 7 (McNemar p < 0.05). We conclude that Pref(s) explains much of the in-distribution gain, and that computing the reward from activations appears to help generalization.

Questions 1 and 4: Validating that LLM outputs embody the strategy; the 500 dialogues and human agreement

The human preference study uses the same 500 dialogues as the validation study. We validated the generated dialogues first, confirming that each response exhibits its intended strategy, and then collected preferences on them from a separate set of annotators. In the preference study, agreement among the three annotators on the preferred strategy was moderate (Fleiss’ κ = 0.41), as expected when several strategies are acceptable in a state (Table 1). Our analysis relies on the distribution of choices and on membership in the acceptable set (90.4%, 82.0% and 93.8% of choices in Established, Uncertain and Revised), and not on agreement about a single strategy. In the validation, agreement between the majority human label and the LLM jury was κ = 0.78 for acceptability and 0.71 for strategy labels (Table 13).

Questions 2 and 3: “Natural separation” between acceptable and unacceptable strategies, and significance of the best split

Table 1 in the paper shows two groups in each state: strategies in the preferred set receive comparable shares, and every strategy outside it receives a markedly lower percentage share. The preferred sets account for 90.4% (Established), 82.0% (Uncertain) and 93.8% (Revised) of choices, and the smallest share inside each set exceeds the largest share outside it by 86.4, 17.1 and 13.8 percentage points, respectively. We identify these sets by exhaustive search over all 15 two-group splits, scoring each by χ² lack-of-fit (Appendix A.6). To test whether the selected split is reliably better than the runner-up, we bootstrapped the 500 dialogues (10,000 resamples) and repeated the search: the selected split ranks first in 100% (Established), 100% (Uncertain) and 97% (Revised) of 10K resamples. We will revise Section 2.2 to report the coverage, the gaps and this stability analysis directly.

Question 5: Models used in the LLM jury and rubric scorer

The LLM jury consists of Claude Opus 4.6, Gemini 3 Flash and GPT-5.5, using majority vote with GPT-5.5 breaking ties (Figure 7, Table 9, Appendix A.4). The rubric score (R) and user acceptability (UA) are produced by the user simulator, Claude Sonnet 4.6.

Question 6: Rubric dimensions and circularity with user acceptability

The acceptable strategy sets behind SC come from the human preference study, while UA is a persona-conditioned judgment of each individual response by the user simulator. These are separate signals, but both reflect user preferences, so the SC–UA correlation is supporting evidence and not independent validation. Our central finding, the collapse in strategy calibration, does not depend on R or UA.

We hope these responses address your concerns. If they do, we would appreciate it if you would consider updating your assessment of the paper. We are happy to answer any further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.