PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolving Support Priorities in Empathetic Reinforcement Learning

Pengyu Huang, Zhiyuan Han, Wenwen Tong, Hewei Guo, Jiangnan Chen, Sirui Chen, Lewei Lu, Beier Zhu, Xun Yang

arXiv:2609.34249v1cs.AI

TL;DR

Empathetic reinforcement learning often relies on fixed reward rubrics even as support priorities shift during dialogue. CARE adapts rubric priorities turn by turn and achieves state-of-the-art performance across three benchmarks, including higher EMPA EPM-Idx under all three judges.

  • Problem

    Empathetic RL commonly keeps rubric reward priorities fixed across turns, despite support needs changing with dialogue context and effective support depending on current needs.

    Source: §1 · p. 1

  • Method

    CARE generates turn-specific rubrics that adjust the weights and fine-grained criteria of cognitive, affective, and proactive empathy to dialogue context.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    CARE establishes state-of-the-art performance across SentientBench, EQBench3, and EMPA; on EMPA, Gemini-2.5-Pro EPM-Idx rises from 28.11 to 83.54 versus strongest baseline MICA.

    Source: §1 · p. 3, §4.2 · p. 7

  • Takeaways & Limitations

    Learned rubric priorities vary across dialogue stages and user states, showing that rewarded support dimensions adapt as context evolves.

    Source: §1 · p. 3

    CARE is not a mental-health professional or substitute for clinical care or emergency services; high-risk content requires escalation policies, privacy protections, and human oversight.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

We identify a fundamental mismatch in empathetic reinforcement learning: support priorities evolve with the dialogue state, yet existing methods typically optimize predefined reward specifications that remain fixed across turns. To model these evolving support priorities, we organize empathetic support along cognitive, affective, and proactive empathy, and propose Context-Adaptive Rubric Evolution (CARE). At each turn, CARE generates a context-adaptive rubric by adjusting both the weights of these three empathy dimensions and their fine-grained evaluation criteria. The rubric generator is trained with turn-level rubric supervision and human preference data through supervised fine-tuning followed by preference-based reinforcement learning, and then serves as an adaptive reward interface for online empathetic RL. Integrated with both RLVER and MICA, CARE achieves state-of-the-art performance across SentientBench, EQBench3, and EMPA under three independent LLM judges. Notably, on EMPA, CARE improves EPM-Idx over the strongest baseline by at least 13 points under all three judges, including an increase from 28.11 to 83.54 under Gemini-2.5-Pro. Further analyses show that learned rubric priorities systematically vary across dialogue stages and user emotions, demonstrating that CARE adapts what is rewarded as support needs evolve.

1 INTRODUCTION

Because empathetic response quality is open-ended and support needs change during dialogue, fixed reward priorities can mismatch the user’s current needs. CARE addresses this by generating turn-specific rubrics that adapt empathy-dimension weights and evaluation criteria.

1 INTRODUCTION

  • The Principle of Evolving Support Priorities holds that reward priorities should adapt to dialogue state because users’ support needs change as conversations unfold.A user may first need reassurance, later understanding, and eventually actionable guidance.Source: §1 · pp. 1–2
  • CARE(R) raises EMPA EPM-Idx from 28.11 to 83.54 under Gemini-2.5-Pro, and CARE(R) and CARE(M) rank first and second on every reported aggregate metric under all three judges.Across SentientBench, EQBench3, and EMPA, CARE establishes a new state of the art under DeepSeek-V4-Pro, Gemini-2.5-Pro, and GPT-5.5; rubric priorities also shift across dialogue stages and user emotions.Source: Figure 1 · p. 2, §1 · p. 2
  • CARE generates a context-adaptive rubric at each turn, adjusting weights and fine-grained criteria for cognitive, affective, and proactive empathy.Cognitive empathy addresses understanding and reframing concerns, affective empathy recognizes and validates emotions, and proactive empathy strengthens agency and ability to move forward.Source: §1 · p. 2, Figure 1 · p. 2
  • CARE’s rubric generator learns from turn-level rubric supervision and human preference pairs through supervised fine-tuning followed by preference-based reinforcement learning.Supervised fine-tuning learns rubric structure and context-dependent priorities; preference-based reinforcement learning favors rubrics that distinguish human-preferred from dispreferred responses.Source: §1 · p. 2
  • CARE supplies an adaptive reward interface for online empathetic RL that integrates with RLVER and MICA without changing their rollout or optimization procedures.At each turn, a rubric-conditioned evaluator scores candidate responses using a rubric generated from the dialogue context.Source: §1 · p. 2

2 RELATED WORK

Existing empathetic RL methods generally keep reward specifications fixed even as support priorities change across turns. CARE adapts the reward specification to the dialogue state, while rubric-learning work has adapted criteria to other sources of variation.

Empathetic dialogue

Rubric-based reward modeling

  • Rubric-learning work adapts evaluation criteria to queries, user histories, or candidate responses, while CARE adapts them to evolving dialogue state.Rubrics as Rewards (Gunjal et al., 2026) uses natural-language rubrics beyond verifiable domains; query-specific rubric learning (Lv et al., 2026) aligns rubrics with preferences, PARL (Qiu et al., 2026) personalizes them from user histories, and DynamicRubric (Wang et al., 2026a) conditions criteria on current candidate response sets.Source: §2 · p. 3

3 METHOD

CARE learns rubrics that adapt empathy priorities and evaluation criteria to each dialogue turn, then uses them as rewards for online dialogue reinforcement learning.

3.1 SETUP

  • CARE keeps cognitive, affective, and proactive empathy dimensions fixed while adapting their priority weights and fine-grained criteria to each dialogue context.Support priorities shift with dialogue needs: affective empathy may lead early, cognitive empathy as users articulate conflicts, and proactive empathy when they consider next steps.Source: §3.1 · pp. 3–4

3.2 TURN-LEVEL RUBRIC DATA CONSTRUCTION

  • Turn-level rubric targets use MINT and ESConv contexts, reference responses, and mapped support-strategy annotations; two judges filter the generated targets into 1.5K SFT examples.The strategy annotations provide coarse supervision for turn-specific priorities across the three empathy dimensions.Source: §3.2 · p. 4
  • Human preference data comprise 2.8K majority-voted pairs from randomized comparisons of anonymized model responses, with preferred-response strategies mapped to empathy dimensions for priority alignment.The preferred-response strategy counts provide a coarse signal of support priorities for generator training.Source: §3.2 · pp. 4–5

3.3 LEARNING THE RUBRIC GENERATOR

  • The generator is trained with SFT followed by preference RL, combining judge-based response comparisons with a priority-alignment reward that ranks dimension weights from support-strategy counts, and optimized using GRPO.The priority-alignment reward supervises weight ranking rather than exact weight values.Source: §3.3 · p. 5
  • A binary format reward additionally encourages generated rubrics to satisfy the required format.The format reward is 1 for a rubric that meets the requirement and 0 otherwise.Source: §3.3 · p. 5

3.4 ONLINE RL WITH EVOLVING RUBRICS

  • In online RL, the context-generated rubric scores candidate responses and replaces the original reward in RLVER and MICA; each rollout group shares one rubric, with a new one generated as context changes.The evaluator assigns normalized scores for each rubric criterion, which are used to compute the turn-specific policy reward.Source: §3.4 · p. 5

4 EXPERIMENTS

CARE achieves state-of-the-art performance across the evaluated settings, and its ablations show that adaptive dimension weights, turn-specific criteria, and preference-based rubric-generator training contribute to its effectiveness.

4.2 MAIN RESULTS

  • CARE achieves state-of-the-art performance in every evaluated setting and improves both RL pipelines; on EMPA, CARE(R) raises EPM-Idx from 28.11 to 83.54 under Gemini-2.5-Pro.Under DeepSeek-V4-Pro, CARE(R) raises EMPA EPM-Idx from 52.10 to 95.77. Table 1 also reports comparisons across SentientBench, EQBench3, and EMPA under three judges.Source: Table 1 · p. 6, §4.2 · p. 7
  • CARE shows no systematic degradation in coding, mathematical reasoning, or instruction following, while learned rubric weights and criteria vary with dialogue turns and user emotions.The analyses visualize turn-wise dimension weights and the distribution of the largest rubric dimension conditioned on user emotion.Source: §4 · p. 6, Figure 3 · p. 8, Figure 4 · p. 8

4.1 SETUP

  • Experiments compare CARE(R) and CARE(M) with prior methods using a shared Qwen2.5-7B-Instruct backbone, and evaluate CARE on three empathy benchmarks with three independent judges.CARE replaces the original rewards in RLVER and MICA with adaptive rubric-based rewards; a frozen Qwen3-8B rubric generator and DeepSeek-V4-Flash evaluator support reward computation.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

  • On EQBench3, 32B CARE(R) scores 76.35 versus GLM-5.2’s 74.90, while 14B CARE(R) scores 73.98 versus 32B Base’s 73.50.CARE consistently improves EQBench3 Overall and all seven other metrics across Qwen3 scales; its 32B result nearly matches Qwen3.7-Max at 76.60.Source: Table 2 · p. 7, §4.2 · p. 8

4.3 ABLATION STUDIES

  • Removing either adaptive dimension weights or turn-specific rubric criteria degrades SentientBench performance; the full CARE rubric performs best across all metrics.The ablations compare uniform weights, fixed criteria, an MICA reward model, and an SFT-only generator against the full rubric.Source: Table 3 · p. 7, §4.3 · p. 8
  • On 166 held-out preference pairs, preference-based RL raises the chosen-response win rate from SFT’s 47.6% to 59.4%; removing priority alignment lowers it to 56.4%.The rubric generator receives dialogue context, and DeepSeek-V4-Flash scores both responses under its generated rubric.Source: Table 4 · p. 7, §4.3 · p. 8

4.2 MAIN RESULTS

  • Under GPT-5.5 on SentientBench, CARE(R) achieves 30.33% success versus MICA’s 5.67%, while reducing the failure rate to 0%.CARE(R) matches or slightly outperforms CARE(M) on most metrics, largely closing RLVER’s original gap with MICA.Source: §4.2 · p. 7
  • On EQBench3, the best CARE variant improves Overall by 2.55–7.67 points over the strongest baseline across three judges, with CARE(R) and CARE(M) ranking first and second.Human evaluators selected CARE(R) in 51.5% of comparisons, versus 23.9% for MICA and 24.6% for RLVER.Source: §4.2 · pp. 7–8

5 ANALYSES

CARE analyses show that preference learning makes rubric priorities vary across dialogue turns and user emotions, reshaping downstream support behavior while broadly preserving general capabilities. The learned empathy weights also align with distinct support needs in representative cases.

Qualitative cases

  • Representative cases show CARE prioritizing cognitive empathy for relationship-conflict interpretation, affective empathy for self-blame and validation, and proactive empathy for practical help.These examples link learned dimension weights to apparent support needs in different dialogue states.Source: §5 · p. 8

Downstream support behavior

  • CARE changes downstream support behavior: MICA broadens beyond paraphrasing, while RLVER shifts away from paraphrasing and advice toward reappraisal.The analyses identify policy-specific shifts rather than convergence on a fixed support strategy.Source: Figure 5 · p. 9, §5 · p. 9

General capabilities

  • CARE improves LiveCodeBench by an average 5.83 points across four backbones, leaves average MATH-500 essentially unchanged, and produces mixed IFBench changes across model scales.IFBench improves on Qwen2.5-7B and Qwen3-32B and decreases on Qwen3-8B and Qwen3-14B.Source: Table 5 · p. 9, §5 · p. 9

Preference learning and rubric priorities

  • Preference-RL shifts rubric priorities from affective toward cognitive empathy across turns, with the two weights crossing around turn 8; across emotions, priorities shift from cognitive to affective empathy as valence becomes more negative.SFT rubrics remain consistently cognitive-dominant across turns and almost uniformly cognitive-dominant across emotions; proactive empathy remains a minor supplementary component in the emotion analysis.Source: §5 · p. 9

Rubric priorities by emotion

  • Preference-RL rubrics also show systematic variation in downstream strategy distributions across user emotions.This observation is reported in Appendix Figure 10.Source: §5 · p. 9

6 CONCLUSION

CARE addresses evolving support priorities by generating turn-specific empathy rubrics and using them as an adaptive reward interface for online dialogue reinforcement learning. Integrated with RLVER and MICA, it improves benchmark performance, while its use requires safeguards for privacy, oversight, and appropriate clinical escalation.

6 CONCLUSION

  • CARE improves performance with both RLVER and MICA across SentientBench, EQBench3, and EMPA under three LLM judges, with particularly large gains on long-horizon interactive benchmarks.Analyses also find that learned rubric priorities vary systematically across dialogue stages and user emotions and reshape downstream support strategies.Source: §6 · p. 9

ETHICS STATEMENT

  • Sensitive disclosures require privacy protections and consent-aware data practices, while high-risk deployment needs escalation policies and human oversight.Multiple judges reduce dependence on one evaluator but do not eliminate cultural-bias risks; the system is not a substitute for clinical care or emergency services, and monitoring should address manipulation, over-reliance, and advice beyond its competence.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The release will support reproduction with prompts, mappings, training configurations, random seeds, evaluation scripts, and licensed model checkpoints or adapters.The paper also directs readers to sections for rubric construction, training, evaluation, implementation details, component-level results, and qualitative outputs.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

  • Generative AI supported manuscript editing and research methodology, including synthetic data generation and online reinforcement-learning evaluation and rewards.The authors reviewed and verified generated content and take responsibility for the final manuscript; named models include DeepSeek-v4-pro, GPT-5, Gemini-3-Pro-Preview, DeepSeek-V3.2, and DeepSeek-V4-Flash.Source: AI USE STATEMENT · p. 10

A IMPLEMENTATION DETAILS

The implementation uses a shared policy initialization and a frozen, history-conditioned rubric generator and evaluator to provide adaptive rewards during online policy optimization.

A IMPLEMENTATION DETAILS

  • All dialogue policies start from Qwen2.5-7B-Instruct, while a fully fine-tuned Qwen3-8B generator produces verified, history-conditioned rubrics scored by a frozen DeepSeek-V4-Flash evaluator.During policy optimization, the generator and evaluator are frozen; one rubric is shared across each GRPO group, and its score replaces the original RLVER or MICA reward. Training uses batch size 8, learning rate 2 × 10−6, 100 steps, 8-turn dialogues, sampling temperature 1, 6 GRPO rollouts per state, and 8 H100 GPUs; rubric-generator reward parameters are τ = 0.7, λ = 0.1, and γ = 0.3.Source: A IMPLEMENTATION DETAILS · p. 13

B BENCHMARK DETAILS … D RUBRIC-DOMINANCE ANALYSIS

The section describes benchmarks for evaluating emotional-support and general capabilities, gives rubric examples with different empathy priorities, and reports that preference-RL dominance shifts from affective early to cognitive later in dialogue. These passages show how evaluation scope and rubric priorities vary across tasks and dialogue contexts.

B.1 SENTIENTBENCH.; B.2 EQBENCH3.; B.3 EMPA.

  • SentientBench (Zhang et al., 2025), EQBench3, and EMPA (Zhang et al., 2026b) evaluate evolving user states, emotional intelligence, and trajectory-level support outcomes.SentientBench uses interactive simulated users; EQBench3 includes role-play and transcript-analysis tasks with an aggregate Overall score; EMPA measures outcome, efficiency, and stability, summarized by EPM-Idx.Source: §B.1 · p. 13, §B.2 · p. 13, §B.3 · p. 13

B.4 LIVECODEBENCH.; B.5 MATH-500.; B.6 IFBENCH.

C.1 COGNITIVE-EMPATHY-DOMINANT CASE

  • Figure 6 identifies the example as a cognitive-dominant rubric case, indicating that cognition is the rubric’s dominant dimension in this case.Source: Figure 6 · p. 14

C.2 AFFECTIVE-EMPATHY-DOMINANT CASE

  • 0.45, 0.30, and 0.25 are the affective, cognitive, and proactive weights in a case centered on self-blame, validation, and emotional safety.The response challenges the parent’s self-blaming interpretation while preserving the son’s autonomy; criterion scores are 5 and 4 for cognitive, and 5 and 5 for both affective and proactive empathy.Source: Figure 7 · p. 14, §C.2 · p. 15

C.3 PROACTIVE-EMPATHY-DOMINANT CASE

  • 0.45, 0.30, and 0.25 are the proactive, cognitive, and affective weights in a leaking-pipe case emphasizing concrete support and user agency.The response proposes a practical path forward; criterion scores are 4 and 3 for cognitive, 3 and 2 for affective, and 4 and 3 for proactive empathy, reflecting comparatively weak emotional validation.Source: Figure 8 · p. 15, §C.3 · p. 15

D RUBRIC-DOMINANCE ANALYSIS

  • Preference RL shifts from affective dominance in early turns toward cognitive dominance later, while SFT remains almost uniformly cognitive-dominant.The analysis measures which dimension has the strict maximum at each turn; Figure 9 aggregates repeated rubric rollouts by turn before reporting proportions across dialogues.Source: D RUBRIC-DOMINANCE ANALYSIS · p. 15, Figure 9 · p. 16

E SUPPORT-STRATEGY ANALYSIS

Support tactics vary by emotion and method: MICA relies mainly on paraphrasing, RLVER uses a broader, more emotion-sensitive set of tactics, and their adapted models diversify support further.

E SUPPORT-STRATEGY ANALYSIS

  • MICA- and RLVER-based models diversify tactics across appraisal, emotional expression, empowerment, and other behaviors, while MICA is paraphrasing-dominated and RLVER uses more advice, validation, and assistance.The comparison covers emotion-conditioned tactic use across scenario-turns and emotion labels.Source: Figure 10 · p. 16, E SUPPORT-STRATEGY ANALYSIS · p. 16

F ADDITIONAL RESULTS

CARE improves overall results on both EQBench3 and EMPA under all three judges. On EMPA, it achieves its strongest reported gains in Outcome, while also improving Efficiency and Stability over reproduced MICA.

F ADDITIONAL RESULTS

  • EQBench3 improvements are most consistent in DoI, ER, HL, and PEI, while DE also improves under DeepSeek and Gemini; WRM and SD are less stable across judges.The pattern suggests CARE strengthens the substance, targeting, and practicality of emotional support rather than surface style alone.Source: F ADDITIONAL RESULTS · pp. 16–17
  • CARE achieves the best EQBench3 Overall score under all three judges: CARE(M) reaches 57.53 under DeepSeek-V4-Pro, while CARE(R) reaches 59.60 under Gemini-2.5-Pro and 37.65 under GPT-5.5.These correspond to gains of 6.06, 7.67, and 2.55 over reproduced MICA, respectively.Source: F ADDITIONAL RESULTS · p. 16, Table 6 · p. 17
  • 43.67, 55.43, and 13.02 are CARE(R)’s EPM-Idx gains over reproduced MICA under DeepSeek-V4-Pro, Gemini-2.5-Pro, and GPT-5.5, respectively.CARE(R) scores 95.77, 83.54, and 38.44 on EPM-Idx under those judges. EMPA aggregates Outcome, Efficiency, and Stability, with all indices normalized so higher is better; CARE’s largest improvements are in Outcome, particularly RDI, Etot, and Snet, and it also improves Efficiency and Stability over MICA.Source: Table 7 · pp. 17–18, F ADDITIONAL RESULTS · pp. 17–18

G HUMAN PREFERENCE EVALUATION · H STRATEGY-TO-DIMENSION MAPPING

A blinded human preference study found that CARE(R) responses received the largest share of non-tie selections. The paper also maps ESConv and MINT strategy annotations to cognitive, affective, and proactive empathy dimensions.

G HUMAN PREFERENCE EVALUATION

  • The study sampled 50 dialogue contexts from the 166 held-out rubric-generator evaluation examples and compared CARE(R), MICA, and RLVER responses to each final user message.Original chosen and rejected responses were discarded; each model independently generated a response from the same context.Source: G HUMAN PREFERENCE EVALUATION · p. 18
  • Fifteen annotators with psychology backgrounds each received 30 randomly assigned contexts and chose among three anonymized responses, with ties allowed.Model identities were hidden and response order shuffled.Source: G HUMAN PREFERENCE EVALUATION · p. 18
  • CARE(R) received 51.5% of non-tie human selections, versus 23.9% for MICA and 24.6% for RLVER.The rate is the fraction of selections received by each model after tie judgments were excluded.Source: G HUMAN PREFERENCE EVALUATION · p. 18
  • The human preference rate counts each model’s share of selections after tie judgments are excluded.This reports preference among judgments with a clear winner, rather than all judgments.Source: G HUMAN PREFERENCE EVALUATION · p. 18

H STRATEGY-TO-DIMENSION MAPPING

  • Strategies are categorized by primary function: interpreting the seeker’s situation maps to cognitive empathy, emotional attunement or validation to affective empathy, and actionable guidance to proactive empathy.The mapping harmonizes ESConv and MINT strategy annotations into the rubric framework’s three dimensions.Source: H STRATEGY-TO-DIMENSION MAPPING · p. 19
  • Table 9 summarizes the mappings from ESConv and MINT strategies to the three empathy dimensions.Source: Table 9 · p. 19, H STRATEGY-TO-DIMENSION MAPPING · p. 19

I SFT DATA CONSTRUCTION

This section describes the construction of CARE’s SFT corpus from emotion-related MINT and ESConv dialogues, including turn-specific rubrics mapped to its three empathy dimensions and screened for quality. It also illustrates how a rubric assigns dimension weights and evaluation criteria for a specific seeker situation.

I SFT DATA CONSTRUCTION

  • Rubrics are generated turn by turn from filtered emotion-related MINT and ESConv dialogues, using mapped empathy strategies and the dialogue history, current query, and reference response.DeepSeek-v4-pro generates the rubrics after the original strategies are mapped to CARE’s three empathy dimensions.Source: I SFT DATA CONSTRUCTION · p. 19
  • Two judges accept a rubric only when all three categories pass turn-specificity, clarity, and non-redundancy checks and their aggregated response scores differ by no more than 0.2 points.Judges score the original supporter response from 0 to 5 on each criterion before category-weighted aggregation.Source: I SFT DATA CONSTRUCTION · p. 19
  • The final SFT corpus contains 1.5k cleaned examples averaging 8.63 dialogue turns, with rubrics averaging 7.67 criteria and 970.72 words.Each criterion description averages 102.72 words.Source: I SFT DATA CONSTRUCTION · p. 19
  • The example rubric weights Cognitive Empathy at 0.4, Proactive Empathy at 0.35, and Affective Empathy at 0.25.Its rationales prioritize understanding the multifaceted conflict and offering actionable support, while recognizing the value of validating the seeker’s and team member’s emotions.Source: I SFT DATA CONSTRUCTION · pp. 26, 28
  • Cognitive criteria assess how well a response connects the team member’s resistance with the need for operational symmetry and offers a coherent conversational strategy.The rubric rewards understanding the resistant member’s perspective and sequencing inquiry with benefits tailored to their concerns.Source: I SFT DATA CONSTRUCTION · p. 27
  • Affective criteria evaluate validation of the seeker’s compounded challenge and recognition that the team member may feel discomfort, frustration, or threatened ownership.The rubric frames resistance as potentially understandable when an existing process works and change may feel top-down.Source: I SFT DATA CONSTRUCTION · p. 28
  • Proactive criteria reward concrete conversation tactics, affirmation of the seeker’s leadership, and guidance for pacing change and ongoing support.Higher ratings call for tailored phrasing and collaborative problem-solving, empowerment tied to the seeker’s advocacy, and phased change with feedback or checkpoints.Source: I SFT DATA CONSTRUCTION · p. 29

J PREFERENCE DATA CONSTRUCTION · K PROMPTS

The preference-data section describes how turn-matched response comparisons are collected and labeled for rubric-generator preference learning. The prompts specify turn-adaptive rubrics across three empathy dimensions, with validation and formatting constraints for training and inference.

J PREFERENCE DATA CONSTRUCTION

  • The rubric-generator preference set contains 2.8k pairs, comparing responses from GPT-5, Gemini-3-Pro-Preview, and DeepSeek-V3.2 generated from the same preceding context.Each pair is anonymized and randomized, then judged by three annotators; the majority decision supplies the preference label. Examples average 12.42 turns, with seeker turns averaging 31.04 words and chosen and rejected supporter responses averaging 75.78 and 70.00 words, respectively.Source: J PREFERENCE DATA CONSTRUCTION · pp. 29–30
  • For majority-preferred responses, coarse support-strategy labels from the MINT annotator are mapped to empathy dimensions using the fixed mapping in § H.Source: J PREFERENCE DATA CONSTRUCTION · p. 30

K PROMPTS

  • The SFT tagger prompt uses conversation context, calibration response, tactic signals, schema example, and validator feedback to generate analysis fields and the rubric.Source: K PROMPTS · p. 32
  • The rubric specifies a weight rationale and category weight for each category, requires 2–4 criteria per category, and sets category weights to sum to 1.0.Criteria counts should reflect the turn’s tactic mix and support needs; rubric text should avoid prompt/meta wording and calibration-response leakage.Source: K PROMPTS · pp. 32–33
  • The SFT workflow repairs validator-identified schema, criteria-count, level-format, title, category, or weight-sum errors and preserves locked categories during regeneration.Its final self-check covers JSON shape, calibration-phrase copying, forbidden prompt words, rationale ranks, and tactic coverage.Source: K PROMPTS · p. 33
  • The validation judge checks category clarity, non-redundancy, and turn-specific fit, and scores each rubric criterion from 0 to 5 without producing a total.Source: K PROMPTS · p. 33
  • At inference, the rubric-generator prompt requests JSON-only output with category weights, concise weight rationales tied to the latest seeker need and supporter tactics, and criteria.Source: K PROMPTS · p. 34

Related papers

Loading 2609.34249v1…