PaperlayerView exact revision on arXiv

Source-linked AI summary

CLUE: Uncovering Social Context for Pragmatic Understanding

Chengfei Wu, Zhaoqing Wu, Dan Goldwasser

arXiv:2610.22167v1cs.CL

TL;DR

Models can benefit from social context, but relevant cues may already lie across a conversation and still be overlooked. CLUE trains models to connect quoted evidence to predictions and rewards new, grounded inferences. It improves average weighted F1 in- and out-of-domain while reducing reasoning-token use.

  • Problem

    Models can overlook earlier conversational exchanges that change an interpretation, even when relevant social context is present in the interaction.

    Source: §1 · p. 1

  • Method

    CLUE uses supervised evidence reasoning followed by GRPO guided by IUG, which rewards new, evidence-supported social inferences while accounting for reasoning effort.

    Source: §1 · p. 2

  • Results

    17.3% in-domain and 9.3% out-of-domain relative gains in average weighted F1 over base Qwen, with over 80% fewer reasoning tokens on average in both groups.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings support context discovery as a learning objective for social and pragmatic understanding.

    Source: §6 · p. 9

    Evaluation covers Chinese, Japanese, and English in selected conversational settings, and computational constraints limited training to a 9B model.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Understanding a conversation requires connecting what is said to the speakers' relationships, expectations, and earlier exchanges. Prior work shows that supplying relevant context can improve models' social judgments, but much of this information may already be present in the conversation. This motivates training models to discover the cues that matter and use them to support an interpretation. We propose CLUE, a two-stage training pipeline that combines supervised evidence reasoning with GRPO guided by Grounded Information Utilization ($\mathrm{IU}_{\mathrm{G}}$). This reward credits new, evidence-supported social inferences while accounting for reasoning effort. We train on a small set of conversational examples to assess effectiveness and generalization across tasks, languages, and genres. At inference time, the trained model reasons over the supplied interaction in a self-contained manner. Compared with base Qwen, CLUE achieves relative gains of 17.3% in-domain and 9.3% out-of-domain in average weighted F1, with higher F1 on four of five out-of-domain evaluations. These gains are accompanied by over 80% fewer reasoning tokens on average in both domains. Human evaluation further finds that CLUE more often grounds its explanations in specific conversational content, with closer alignment between its quoted evidence and human-selected context.

1 Introduction

CLUE trains models to find and use conversational cues already present in an interaction, addressing the difficulty of deciding which earlier exchanges inform an interpretation. It improves average weighted F1 in- and out-of-domain while using fewer reasoning tokens on average.

1 Introduction

  • An earlier playful use of an expression makes its later repetition in response to an extravagant gift indicate surprise rather than anger.The example illustrates why an exclamation can be ambiguous when read in isolation.Source: §1 · p. 1
  • Prior work found that supplying cultural norms or context schemas can improve social judgments, motivating models to discover relevant context already present in conversations.Relevant evidence in long interactions may be spread across turns, allowing plausible explanations to overlook exchanges that change an interpretation.Source: §1 · p. 1
  • CLUE combines supervised evidence reasoning with GRPO guided by IUG, rewarding new, evidence-supported social inferences while accounting for reasoning effort.At inference, the trained model reasons over the supplied interaction in a self-contained manner.Source: §1 · p. 2
  • 17.3% in-domain and 9.3% out-of-domain relative gains in average weighted F1 over base Qwen, with higher F1 on four of five out-of-domain evaluations.The gains accompanied over 80% fewer reasoning tokens on average in both groups.Source: §1 · p. 2
  • Human evaluation found that CLUE more often links specific conversational content to predictions, and its quoted evidence better aligned with lines selected by human readers.Analysis of an emotion-cause dataset also found that CLUE more often identifies causes in its reasoning, which may help explain improved emotion recognition.Source: §1 · p. 2

2 Related Work

Related work covers conversational social context, social-norm resources, and verifiable rewards. CLUE applies evidence-linked reasoning and a reward for new, grounded social inferences to context discovery.

2 Related Work

  • Prior work studies explicit norms, extracted social knowledge, conversational power prediction, political discourse, and reasoning grounded in visual evidence.NormAd (Rao et al., 2025) evaluates reasoning with explicit norms; NormSage (Fung et al., 2023) and cultural context schemas extract or organize conversational social knowledge.Source: §2 · p. 2
  • Social-norm resources encode rules, narratives, and situational constraints, while SocialDial and NormDial study norm observance in dialogue.The paper uses CCU conversations for human annotations across emotion, dialog act, and norm violation, supporting context discovery with verifiable feedback.Source: §2 · p. 2

Reinforcement learning with verifiable rewards

  • Unlike rewards for retrieval, intermediate-step feedback, or evidence-supported facts, IUG rewards new evidence-supported social inferences while accounting for reasoning effort.CLUE initializes with supervised reasoning traces connecting quoted conversational evidence to task predictions, then optimizes with GRPO and task correctness.Source: Reinforcement learning with verifiable… · p. 2, Figure 2 · p. 3

3 Method

CLUE trains a policy to connect conversational evidence to task predictions, then uses GRPO to reward correct answers and novel, grounded inferences while accounting for reasoning effort.

3 Method

  • CLUE uses supervised evidence reasoning followed by GRPO with task correctness and Grounded Information Utilization to train social and pragmatic understanding.The reward values new, evidence-supported inferences while accounting for reasoning effort.Source: §3 · p. 3

3.1 Setting and Supervised Initialization

  • The policy takes a conversation, prediction target, and task inputs, then produces reasoning blocks and predicted labels with quoted source spans marked by evidence tags.Supervised initialization refines model-sampled traces using human-guided examples and trains with a weighted completion loss.Source: §3.1 · p. 3

3.2 Rewarding Context Discovery

  • The IUG judge credits a block only for a new, task-relevant inference supported by verified conversational quotations; repetition and unsupported label elimination earn no novelty credit.A quotation verifier checks that quoted text matches the supplied conversation, while the judge assesses semantic support and novelty.Source: §3.2 · pp. 3–4
  • IUG combines inference credit with effective reasoning effort, which subtracts a bounded allowance for verified quote tokens; the combined reward also includes task correctness and evidence format.Reward weights are chosen so that an exact answer receives more total reward than any inexact answer for the same prompt and training step.Source: §3.2 · pp. 4–5

3.3 Advantage and Policy Loss

  • GRPO optimizes group-relative advantages from sampled completions using the combined reward, with a length-normalized policy loss and masking for truncated completions or unusable groups.At inference, the trained policy generates reasoning and a prediction from the supplied interaction without the judge.Source: §3.3 · p. 5

4 Experiments

The experiments evaluate CLUE against supervised and RL baselines across social and pragmatic tasks, then examine efficiency, out-of-domain transfer, ablations, and general capability retention. CLUE improves average weighted F1 over base Qwen in both domains while reducing reasoning-token use.

4.1 Experimental Settings

  • Training uses 2,568 CCU Mandarin instances; evaluation covers held-out Mandarin and five out-of-domain settings across emotion, dialog act, norm violation, and power prediction.The comparisons include Label SFT, Reasoning SFT, two RL variants, and the larger Qwen3.8-27B reference.Source: §4.1 · p. 5

4.2 Performance of CLUE

  • 17.3% in-domain and 9.3% out-of-domain relative gains in average weighted F1 over base Qwen; CLUE also exceeds Qwen3.8-27B on five of eight tasks.Tasks are weighted equally within each domain.Source: §4.2 · p. 5, Table 1 · p. 6

4.3 Reasoning Efficiency

  • From the same supervised initialization, CLUE achieves higher F1 with fewer reasoning tokens than Accuracy-Only RL in both domains.Figure 5 also shows a sustained later-RL reduction in CLUE’s mean completion length, while Accuracy-Only RL changes more modestly.Source: §4.3 · pp. 6–7, Figure 4 · p. 6, Figure 5 · p. 6

4.4 Generalization

  • Despite training only on CCU Mandarin, CLUE improves over base Qwen and Reasoning SFT on Japanese, MELD, and CPED emotion detection; power prediction remains near base performance.The CCU Japanese weighted-F1 gain over base Qwen is 20.92 points; on MELD, CLUE recovers the supervised-initialization loss and surpasses base Qwen.Source: §4.4 · p. 7, Table 2 · p. 7

4.5 Ablation Study

  • CLUE has higher mean weighted and macro F1 than Accuracy-Only RL on all eight tasks, including 3.62 additional weighted-F1 points on CPED dialog act recognition.Both RL variants share the Reasoning SFT initialization, evidence instructions, data, and optimization schedule.Source: §4.5 · p. 7, Table 3 · p. 7

4.6 General Capability Retention

  • Both RL variants remain at or above the base model’s macro F1 on MMLU, the evaluation used to check general knowledge and reasoning beyond social tasks.MMLU uses greedy decoding.Source: §4.6 · p. 7, Table 4 · p. 7

5 Qualitative Analysis

The qualitative analysis examines whether CLUE identifies earlier exchanges that explain emotion and whether its explanations connect conversational evidence to predictions and human-selected context.

5.1 Identifying Emotional Causes

  • The cause analysis tested 825 non-neutral targets with 1,046 annotated earlier cause–target pairs, using MELD prompts and decoding settings.A frontier judge checked whether reasoning used an annotated earlier turn as a main cause or contributing factor.Source: §5.1 · pp. 7–8
  • CLUE identified annotated causes across more target utterances while also improving emotion recognition, supporting its ability to use earlier exchanges to explain judgments.Figure 6 reports cause identification and weighted emotion F1, averaged over three sampled evaluations; cause identification is the percentage of targets with at least one annotated cause identified in reasoning.Source: §5.1 · p. 8, Figure 6 · p. 8

5.2 Human Evaluation

  • Human evaluation compared paired outputs on 50 instances across eight tasks, with annotators selecting needed lines before rating anonymized reasoning excerpts.Grounding required specific conversational content and an explanation of how it supported the answer; Evidence Alignment F1 measured overlap with human-selected lines.Source: §5.2 · p. 8
  • 85% of CLUE excerpts were rated grounded, compared with 33% for base Qwen; Evidence Alignment F1 was higher overall, with the clearest gain in-domain.On out-of-domain items, grounding remained higher, but the smaller alignment difference was not statistically significant; among 22 instances where both models answered correctly, grounding was 95.5% versus 50.0%.Source: §5.2 · p. 8, Figure 7 · p. 8

6 Conclusion

The conclusion presents CLUE as a context-discovery training pipeline and reports improved F1 and reduced reasoning-token use on limited conversational data. The qualitative findings support context discovery as a learning objective for social and pragmatic understanding.

6 Conclusion

  • CLUE combines supervised evidence reasoning with GRPO guided by IUG, which rewards new evidence-supported social inferences while accounting for reasoning effort.Training on limited conversational data improves average F1 in-domain and out-of-domain while reducing reasoning-token use.Source: §6 · p. 9
  • CLUE more often identifies earlier conversational causes of emotion in its qualitative analysis.Source: §6 · p. 9
  • The findings support context discovery as a learning objective for social and pragmatic understanding.Source: §6 · p. 9

Limitations

The evaluation covers Chinese, Japanese, and English in selected conversational settings, leaving broader cultural coverage to be explored.

Limitations

  • The evaluation covers Chinese, Japanese, and English in selected conversational settings, leaving broader cultural coverage to be explored.Training was limited by computational constraints to a 9B model, with a 27B model included as a reference; broader communities, languages, and scales remain future directions.Source: Limitations · p. 9

A Training Setup

The training setup uses NVIDIA H100 GPUs for both supervised initialization and reinforcement learning.

A Training Setup

  • Supervised initialization uses 1 NVIDIA H100 GPU, while reinforcement learning uses 3 NVIDIA H100 GPUs.Source: A Training Setup · p. 11

A.1 Supervised Initialization

Supervised initialization refines sampled reasoning traces and selects a checkpoint using held-out macro F1 across tasks. Training also specifies optimization, parameter-freezing, and loss-allocation settings.

A.1 Supervised Initialization

  • GPT-6-family models refine sampled reasoning traces, and the stage 1 checkpoint is selected by mean macro F1 across 24 held-out targets, eight per task.Rows sharing those targets are excluded from supervised fine-tuning.Source: §A.1 · p. 11
  • SFT runs for two epochs with AdamW, one example per device, four gradient-accumulation steps, and seed 42.The learning rate warms up over 1% of training and then decays linearly to zero.Source: §A.1 · p. 11
  • Input embeddings and lm_head remain frozen except for the evidence-token rows, while 15% of completion loss weight goes to final-label tokens.The remaining 85% covers other tokens, including the end-of-turn token.Source: §A.1 · p. 11

A.2 Reinforcement Learning

Reinforcement learning uses AdamW and a staged learning-rate schedule, with prompts drawn from progressively varied conversational contexts. CLUE additionally retries unusable groups and uses a detached sampler correction.

A.2 Reinforcement Learning

  • RL trains for 723 updates with eight prompts per update and K = 8 completions per prompt, using AdamW and frozen input embeddings and lm_head.The learning rate warms up over 19 updates, stays constant through update 506, then decays by cosine to its initial value; the final checkpoint is evaluated without selection.Source: §A.2 · p. 11
  • Training moves from full conversations ordered by difficulty to mixed contexts containing the target and its preceding 5, 10, or 20 turns.Accuracy-Only RL uses the same data mix; CLUE retries unusable groups up to three times, then tries up to two replacement prompts.Source: §A.2 · p. 11
  • SFT checkpoint selection and RL rollouts use temperature 1, top-p 1, no top-k truncation, and no separate reasoning budget.RL rollout token limits come from Table 5; the detached sampler correction is introduced for h_i,t = (x, y_i,<t).Source: §A.2 · p. 11
  • The sampler correction settings are ρmax = 3, εstd = 10−4, δ = 1.5, ϵ− = 3 × 10−4, and ϵ+ = 4 × 10−4.Source: §A.2 · p. 12

A.3 Reward Parameters

The reward setup shares hyperparameters across tasks, including an effort budget, and schedules ramps around the transition to mixed context lengths. The judge uses a constrained decoding configuration.

A.3 Reward Parameters

  • All tasks share reward hyperparameters, including the effort budget B.Table 6 is identified as the source of CLUE’s reward parameters.Source: §A.3 · p. 12
  • The reward ramps end at the transition to mixed context lengths, at approximately τ = 0.405.Source: §A.3 · p. 12
  • The judge runs with thinking disabled, temperature 0.3, top-p 0.95, top-k 20, and an output budget of 128 + 120N tokens.Source: §A.3 · p. 12

B Experimental Details

The experiments evaluate social, pragmatic, and MMLU tasks using specified dataset splits and input conventions. Power prediction uses administrator status as a proxy, and answers are scored only when explicitly extractable.

B Experimental Details

  • CCU Mandarin and Japanese use held-out Evaluation partitions, while MELD, CPED, ECF 2.0, and MMLU use test splits.Turn-level inputs end at the target; forum chains are preserved, and messaging gaps of at least 10 minutes appear as time markers without counting as turns.Source: B Experimental Details · p. 12
  • Evaluation set sizes count MMLU questions, power-prediction conversations, and target turns for the other tasks.Source: Table 8 · p. 12
  • Power prediction uses English Wikipedia Talk Pages conversations, with administrator status serving as a proxy for institutional power.The first 1,000 conversations with exactly two speakers and at least one administrator utterance are retained; speakers are labeled by first appearance.Source: B Experimental Details · p. 12
  • Only explicit answer fields or label lists are accepted; empty or invalid answers are incorrect, and catch-all gold-label turns remain context but not prediction targets.The main experiments use GH200-120GB or H100 GPUs, and social and pragmatic tasks use three completions per instance.Source: B Experimental Details · p. 12

B.1 Annotation Interface

Figure 8 shows the annotation interface used to select conversational evidence and rate anonymized reasoning excerpts.

B.1 Annotation Interface

C Prompt Templates

The prompt templates operationalize evidence-grounded social judgments and explanation evaluation or refinement, directing models to use relevant conversational context, distinguish supported contributions from repetition or unsupported claims, and follow constrained formats. They cover tasks including emotion, dialogue-act, power, norm-adherence, and forum-emotion judgments, as well as IUG judging and SFT-generation records.

C.1 Inference

  • Emotion classification targets the specified turn, using other turns only as context and considering conveyed emotion rather than literal wording alone.The template also cautions against confusing emotion with dialog act or norm violation and against inventing unsupported facts or causes.Source: §C.1 · p. 14
  • The transcript guidance treats short segments as lines, allows adjacent messages to be read together when they continue a turn, and treats elapsed-time notes as non-turns.The prompt leaves it to the text whether preceding lines run into the target or the target stands alone.Source: §C.1 · p. 14
  • Reasoning instructions require concrete claims to be grounded in exact conversation text, with verbatim quotations enclosed in <evidence> tags.Relevant context may include relationships, status, conflict, obligations, expectations, and responses to praise, blame, pressure, or reassurance.Source: §C.1 · p. 14
  • The emotion prompt restricts answers to its defined labels, permits multiple clearly present emotions or None, and requires the final answer in English on the last line.The provided options are Anger, Anticipation, Joy, Trust, Fear, Surprise, Sadness, Disgust, and None.Source: §C.1 · p. 14
  • Dialog-act recognition classifies the target turn’s pragmatic intention, grounding claims in exact text and interpreting what the turn does in context rather than literally alone.The template distinguishes dialog act from emotion, politeness, disagreement, and norm violation, and warns against unsupported inferences.Source: §C.1 · pp. 14–15
  • The dialog-act prompt supplies defined act options and requires an English final line containing only one or more applicable labels, or None.Options include apology, criticism, greeting, request, persuasion, thanks, taking leave, admiration, finalizing a negotiation or deal, and refusing a request.Source: §C.1 · p. 15
  • Norm-violation detection judges whether the target turn meets expectations associated with its provided dialog act, using supported context rather than general judgments.The prompt asks models to consider active expectations and distinguish norm adherence from emotion, disagreement, dialog act, or politeness markers.Source: §C.1 · p. 15
  • A separate TV-series template asks for exactly one target-utterance emotion from anger, disgust, fear, joy, neutral, sadness, or surprise.The input is a conversation between characters followed by a target utterance.Source: §C.1 · pp. 15–16
  • Emotion prompts require exact quoted evidence and relevant conversational cues, reject unsupported claims, and classify utterances with the specified emotion labels.One prompt treats an utterance without a particular emotion as neutral.Source: §C.1 · p. 16
  • The Chinese-series emotion template provides label definitions, requires evidence quotes to preserve Chinese text, and requires a final English label from its specified set.The set includes positive, negative, and neutral categories, including happy, grateful, relaxed, anger, sadness, and worried.Source: §C.1 · p. 16
  • The Chinese-series dialogue-act template asks models to infer an utterance’s communicative function from context, not literal wording alone, and to distinguish that function from emotion.It highlights reactions to prior questions, requests, offers, or statements and how others respond.Source: §C.1 · p. 16
  • Dialogue-act prompts provide definitions for the candidate acts and require one exact English label as the final answer.The inventory includes acts such as greeting, question, answer, apology, agreement, irony, comfort, and other.Source: §C.1 · pp. 16–17
  • The power prompt defines standing through authority in the setting and asks models to infer it from text-supported interactional cues.Cues include requests for decisions, compliance with directives, authority-specific actions, policy invocation, and deference; speaking more or proposing a plan is not a reliable cue.Source: §C.1 · p. 17
  • The appendix also begins a Japanese discussion-forum emotion-detection instruction, introducing a forum thread as input.The supplied excerpt ends before giving the full instruction.Source: §C.1 · p. 17
  • The inference prompt classifies the target writer’s emotion in context, while requiring every concrete conversational claim to be supported by exact text enclosed in evidence tags.The target line is judged using the rest of the thread as context; quoted evidence must be copied exactly, not paraphrased.Source: §C.1 · p. 17
  • The prompt asks the model to consider relevant social context but judge the target writer’s conveyed emotion, not merely literal wording or another person’s emotion.Relevant cues include relationships, status, thread tone, conflict, support, expectations, and the target post’s conversational response; unsupported facts and motivations are prohibited.Source: §C.1 · pp. 17–18
  • The output must use the defined emotion labels, with “None” for unclear or unfitting cases, and end with an exact English answer line.The prompt distinguishes emotion from dialog act or norm violation and specifies comma-separated labels when multiple emotions are clearly present.Source: §C.1 · p. 18

C.2 IUG Judge

  • The IUG judge evaluates whether each reasoning paragraph adds a new, task-relevant interpretation supported by quoted evidence, rather than answering the underlying task or assuming a gold answer.It processes paragraphs in order against all earlier supported contributions, with novelty measured relative to warranted earlier interpretations.Source: §C.2 · p. 18
  • A contribution must be verified against the source and task information; an unquoted premise, misleading quotation, or merely repeated interpretation does not earn progress.Evidence can be combined or reused to establish a new insight, but changing wording, labels, or confidence alone does not make an interpretation novel.Source: §C.2 · pp. 18–19
  • The judge rejects restatement and label mapping as progress, but can credit a label comparison that resolves a plausible alternative with a new evidence-backed distinction.A short inference from quoted evidence to a task decision may qualify; exhaustive elimination of implausible alternatives does not.Source: §C.2 · pp. 18–19
  • Specific self-correction can count when it establishes a supported change in interpretation, while general doubt earns no credit and mixed paragraphs qualify if a valid new contribution remains.The corrected interpretation is counted once, without a bonus for the earlier mistake; unrelated errors do not invalidate an independently supported contribution, but a withdrawn or contradicted claim cannot qualify.Source: §C.2 · p. 19
  • The IUG judge credits new, source-supported contributions—not evidence length or agreement with a final label—and does not credit repetition or unsupported claims.Brief spans can support an interpretation, while long quotations may add nothing; mixed paragraphs can receive credit for a valid contribution without endorsing errors.Source: §C.2 · p. 19
  • The judge returns one short English assessment and a GAIN or NO_GAIN verdict per paragraph, in order, using the prescribed one-line format.Assessments name the new contribution or the reason no supported improvement exists; for mixed GAIN paragraphs, they identify the valid contribution rather than endorsing errors.Source: §C.2 · p. 19
  • Independent calibration examples illustrate crediting useful interpretations, supported corrections, and limits while rejecting repetition, general doubt, and unsupported conclusions.The examples distinguish a timing refusal from a later review, correction of speaker attribution from unspecific doubt, and requested redaction from verified compliance.Source: §C.2 · pp. 19–20

C.3 SFT Data Generation

  • The refinement template treats decoded prompt messages as authoritative, preserves the original identifier and task, and requires the output prompt to retain the complete message list unchanged.It also normalizes ground-truth labels to the task’s exact options without changing their meaning or adding, removing, or substituting labels.Source: §C.3 · p. 20
  • Refinement edits the original explanation toward ground-truth labels by adding missing reasoning and corrections while preserving supported reasoning and intelligible mistaken inferences.The template says to keep the original completion as the editing base and not silently replace the correction process with a polished argument.Source: §C.3 · p. 21
  • The editing procedure favors local insertion, preserves supported sentences verbatim, and places new target interpretations after the inference they check rather than in opening background.It also calls for checking intervening turns and recorded time gaps when repairing claims about adjacency or topic continuity, while retaining plausible older causes.Source: §C.3 · p. 21
  • Templates require reading the full task, conversation, target turn, and original completion, treating supplied labels as correction targets rather than conversational evidence.If the conversation and task definitions cannot support the intended labels, the instructions call for a separate conflict procedure rather than an invented premise or silent label change.Source: §C.3 · p. 21
  • Interpret the target through the unfolding exchange, using supported relationship, turn-order, timing, and emotional cues rather than generic context or stereotypes.Templates call for connecting relevant earlier activity or intervening turns to the target while distinguishing likely causes from stated facts and keeping each speaker’s state separate.Source: §C.3 · pp. 21–22
  • Preserve supported original wording and its order, including an intelligible mistaken conclusion, then place a clear correction after the inference it examines.Do not rewrite sound material for style or erase the initial commitment; the subsequent check must retract it, and later reasoning must not reassert the rejected conclusion.Source: §C.3 · p. 22
  • Choose the smallest intervention that addresses the actual defect, using insertion or local revision when possible and reserving rewrites for corrupted or incoherent reasoning.Mechanical transcription errors should be repaired locally, while unsupported labels require a conflict record; brevity or stylistic preference alone does not justify deleting or replacing substantive reasoning.Source: §C.3 · p. 22
  • A substantive correction identifies the precise inference, checks it against the source and task definition, explains the update, independently supports the intended label, and carries the update forward.The observable sequence should be the existing inference, a specific check, and the resulting update; a natural rereading move may help, but no stock phrase is required.Source: §C.3 · p. 22
  • Add only missing, source-supported links—such as context, referents, emotional causes, or label distinctions—and explain how they connect the recent interaction to the target’s wording and label.Templates require explaining what cues such as repetition, hesitation, or punctuation do in context, whose feeling they convey, and how that state carries into or changes at the target; they discourage generic outlines and exhaustive additions.Source: §C.3 · p. 23
  • Explanations should distinguish labels by their defining conditions, address credible alternatives fairly, preserve uncertainty, and explain how multiple labels can coexist.The goal is a sufficient local explanation, not extra length or a forced single dominant label.Source: §C.3 · p. 23
  • Evidence spans must be nonempty, exact contiguous source text, and fabricated quotations or unsupported claims must be repaired or removed.After correction, no conclusion may rely on a rejected assumption, and evidence must not contain paraphrases, translations, or joined fragments.Source: §C.3 · p. 23
  • Interpretations must check attribution, referents, negation, time, modality, and scope; a known software-generated reply is an automated notice, not proof of intent or response.Questions are not confirmations, offers are not completed actions, and reported beliefs remain the speaker’s claims.Source: §C.3 · p. 23
  • For ED and ND, apply the supplied definitions to selections and exclusions, allow supported implicit or multiple labels, and flag unsupported or missing labels as conflicts.A negative event alone does not determine an emotion, and a pragmatic inference must satisfy the act’s defining conditions.Source: §C.3 · p. 23
  • For VD, assess the target against a context-supported expectation and the assigned dialogue act, preserving uncertainty rather than inventing a norm or obligation.Refinements should reach the supplied label set only when it is source-grounded; unsupported targets or changed definitions require a conflict flag.Source: §C.3 · p. 24
  • A successful refinement returns the complete edited explanation in its original language and organization, with one think block and a final normalized Answer line.Retain a meaningful initial mistake when present, insert its source-grounded correction, and continue with the revised understanding.Source: §C.3 · p. 24
  • Successful rows contain unchanged UUID, task, and decoded prompt messages plus exactly one assistant completion, with no labels or audit fields.The prompt and completion are JSON arrays of role/content objects, not JSON-encoded strings.Source: §C.3 · p. 24
  • Unsupported supplied labels require a separate reference-conflict audit record, not a fabricated training answer or a silently substituted alternative.The final checks reject refinements that erase a meaningful mistake or reach the target through unsupported reasoning.Source: §C.3 · pp. 24–25
  • SFT-generation records contain fields for UUID, task, prompt messages, completion text, prompt text, and labels.The supplied record template uses those fields as its JSON packet structure.Source: §C.3 · p. 25

C.4 ECF 2.0 Cause Judge

  • The ECF 2.0 cause judge evaluates whether a specific assessed turn serves as a reason for the emotion in the candidate’s final answer.It receives the task, numbered conversation with the target turn identified, candidate reasoning and answer, and the turn to assess; it judges the turn’s role, not whether the emotion is correct.Source: §C.4 · p. 25
  • The judge returns YES when reasoning connects the assessed turn to the concluded emotion, including as a reaction, trigger, or the speaker’s earlier words expressing that emotion.Any one such role is sufficient, even when the turn is only one of several reasons.Source: §C.4 · p. 25
  • The judge returns NO when a turn is only summarized, used as background or contrast, echoed by the target, or linked only to an emotion the reasoning rejects.Merely naming what the target replies to does not establish that the turn supports the concluded emotion.Source: §C.4 · p. 25
  • The decision depends on the turn’s role in the reasoning, not its label or causal wording; neutral judgments count only when the assessed turn itself supports the calm interpretation.A turn called “context” can receive YES if the reasoning says the speaker responds to it with the concluded emotion; a neutral reading based only on the target’s words or tone receives NO.Source: §C.4 · p. 25
  • A reference counts only when the reasoning identifies the assessed turn clearly; the judge matches misquoted turns by content and evaluates only written reasoning, without supplying missing causal links.The same rules apply across languages, and a quotation alone is not a reason.Source: §C.4 · p. 25
  • The judge must output exactly one line containing a short assessment of at most 20 words, followed by YES or NO, with no headings or extra formatting.The assessment briefly describes how the reasoning uses the turn.Source: §C.4 · p. 25

Related papers

Loading 2610.22167v1…