PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Communication-Efficient Social Intelligence in Language Agents

Linxiao Gong, Yijie Xu, Tianfu Wang, Yin Wu, Yili Wang, Xingbo Yao, Huizai Yao, Xilin Xia, Haowen Yang, Hui Xiong

arXiv:2609.35749v1cs.CL

TL;DR

Social-agent communication must balance goal attainment with the words and exchanges needed across an interaction, since short responses can leave partner constraints unresolved. TACT tests expression and strategy revisions through partner replies, selects guidance by local goal support and token cost, and distills feedback on student prefixes. It achieves the highest Goal on SOTOPIA All and Hard with substantially fewer target tokens than SFT+SDPO, and improves AgentSense goal success while reducing tokens and messages.

  • Problem

    Response length alone does not capture communication needed across an interaction, where action choices and expression shape partner understanding, responses, and subsequent exchanges.

    Source: §1 · p. 1

  • Method

    TACT tests expression and strategy revisions to student-generated actions through partner responses, selects a reference by local goal support and token cost, and distills it on the student’s original prefixes.

    Source: §1 · p. 2

  • Results

    TACT achieves the highest Goal on SOTOPIA All and Hard with substantially fewer target tokens than SFT+SDPO, and improves AgentSense goal success while reducing target tokens and interaction messages.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    Teacher feedback on the student’s original prefixes lets the student act independently at deployment without a specialist.

    Source: §1 · p. 2

    The local proxy can miss ambiguous commitments whose consequences emerge later, so selection does not establish that a candidate is socially sufficient.

    Source: §3.3 · p. 5

Abstract

from arXiv · show

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, we propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding. We first characterize communication efficiency in terms of action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges. We design TACT to revise student-generated actions, test the revisions through partner responses, and distill useful feedback into the student. An expression specialist removes unnecessary detail while preserving the intended action, while a strategy specialist proposes alternatives that may better address the partner's constraints. To determine which revision helps, TACT samples a partner response for each candidate and selects a teacher reference by balancing local goal support against action-token cost. That reference guides on-policy distillation on the student's own generation prefixes, allowing the student to act independently at deployment. We evaluate TACT on SOTOPIA and AgentSense. On SOTOPIA, it achieves the highest Goal among the evaluated methods on All and Hard while using substantially fewer target tokens than SFT+SDPO. On AgentSense, it improves goal success over the initial student while reducing target tokens and interaction messages.

1 INTRODUCTION

Communication efficiency depends on both goal attainment and the communication needed across the interaction, not merely the length of one response. TACT tests expression and strategy revisions through partner responses, selects guidance by local goal support and token cost, and distills it on student-generated prefixes.

1 INTRODUCTION

  • Communication cost depends on goal attainment and communication across the complete interaction, because an action shapes what the partner understands and how they respond.Two interactions can reach the same goal while requiring different amounts of communication.Source: §1 · pp. 1–2
  • Training must turn feedback into concrete action revisions, assess proposals through interaction, balance goal progress against communication cost, and supervise the student’s own generation prefixes.Training only on revised responses can miss the prefixes the student encounters when acting independently.Source: §1 · p. 2
  • On SOTOPIA, TACT achieves the highest Goal on All and Hard with substantially fewer target tokens than SFT+SDPO; on AgentSense, it improves goal success while reducing tokens and messages.The reported ablations show distinct effects of expression and strategy supervision on token use and interaction turns.Source: §1 · pp. 2–3
  • Avoidable communication cost comes from excess wording within an action and further exchanges that follow when the action leaves partner constraints unresolved.An agent must choose both an action that can move the interaction forward and expression suited to what the partner already knows.Source: §1 · pp. 1–2
  • TACT revises student-generated actions with expression and strategy specialists, tests proposals through partner responses, and selects a reference using local goal support and action-token cost.On-policy distillation gives the reference to the teacher, which provides feedback on the student’s original prefixes; the student needs no specialist at deployment.Source: §1 · p. 2

2 RELATED WORK

Related work connects social-agent learning to interactive evaluation, interaction-based supervision, and on-policy distillation. TACT combines these threads by testing candidate actions through partner responses and giving the teacher feedback on student-generated prefixes.

Social Intelligence

  • Social-intelligence research spans interactive benchmarks, social-skill tutoring, and methods for learning agents’ behavior from interaction data.TACT focuses on expression and strategy revisions as training references for improving communication over complete interactions.Source: §2 · p. 3

Interaction-Based Supervision

  • Interaction-based supervision connects feedback to individual actions, including through process feedback, future-interaction utility, information gain, and simulated continuations.TACT samples one partner response for each original or candidate action, then compares local goal support and token cost under the same context.Source: §2 · p. 3

On-Policy Distillation

  • On-policy distillation provides teacher supervision at student-generated prefixes, and TACT supplies interaction-selected revisions only to the teacher so the student acts independently at deployment.Prior work enriches teacher contexts with demonstrations, solutions, environment feedback, and subsequent user messages.Source: §2 · p. 3

3 METHOD

TACT treats communication efficiency as a joint choice of action and expression, measured across goal attainment and the communication needed over a complete interaction. It tests specialist revisions through partner responses and uses a selected reference to guide on-policy distillation.

3 METHOD

  • TACT compares specialist-proposed revisions with the original action through partner responses, then selects a reference to guide on-policy distillation.The selection balances local goal-support gain against action-token cost.Source: §3 · p. 3, Figure 2 · p. 4

3.1 PROBLEM FORMULATION

  • Interaction outcomes combine final goal attainment with action-token and turn costs across a complete trajectory.Action-token cost sums the target agent’s action lengths; turn cost counts environment turns taken by both participants.Source: §3.1 · pp. 3–4

3.2 DUAL-SPECIALIST REVISION

  • Expression edits shorten wording while preserving intent; strategy edits may change intent or action type to address constraints, and candidate branches do not replace later student turns.Expression edits must also preserve facts, commitments, and action type; strategy edits may accept a constraint, offer an arrangement, seek confirmation, or end the interaction.Source: §3.2 · p. 4

3.3 EFFICIENCY-AWARE SELECTION

  • Each distinct legal action receives one partner reply; eligible alternatives follow specialist-specific goal-support and token-cost rules, with selection balancing support against cost.Expression candidates must be shorter; strategy candidates must improve local goal support over the original.Source: §3 · p. 3, §3.3 · p. 5
  • The one-reply comparison is only a local contribution proxy, so it can miss later consequences such as an agreement remaining ambiguous.The method does not continue candidate branches to the end of the dialogue.Source: §3.3 · p. 5

3.4 ON-POLICY DISTILLATION

  • The selected revision conditions the teacher’s evaluation of the student’s original action tokens, while the student keeps its original context and acts independently at deployment.The student collects refreshed interactions after each batch update.Source: §3.4 · p. 6

4 EXPERIMENTS

The experiments compare TACT with training and prompting baselines on SOTOPIA and AgentSense. TACT attains strong social goals across evaluated settings, with communication costs that vary by method and partner.

4.1 EXPERIMENTAL SETUP

  • The evaluation uses SOTOPIA’s 450-setting All panel and 70-setting Hard subset, plus 500 AgentSense instances from 100 two-person templates.The student is Qwen3.5-4B; only the target agent is trained, and the evaluation partner is the initial Qwen3.5-4B.Source: §4.1 · p. 6

4.2 RESULTS

  • TACT achieves the highest Goal on SOTOPIA All and Hard, exceeding SFT+SDPO in Goal while using substantially fewer target tokens.Against vanilla OPD, it achieves higher Goal at similar target-token cost on All, with more turns; concise prompting uses fewer tokens but has lower Goal.Source: §4.2 · pp. 6–7, Table 1 · p. 7
  • Prompted OPD improves All Goal over vanilla OPD (5.151 versus 4.960) at nearly identical token cost, while Hard Goal is unchanged at 3.686.TACT retains higher Goal on both panels, but uses more tokens and turns than Prompted OPD; training exposure and checkpoint selection differ.Source: §4.2 · p. 7
  • Expression-only training uses fewer target tokens, while strategy-only training uses fewer turns; the full framework achieves higher Goal than either specialist alone.The comparisons use the same pool of 1,129 scenario–character configurations.Source: §4.2 · p. 7, Table 2 · p. 8
  • TACT has higher Goal than IG-only, Token-only, and random ranking on both panels; IG-only exceeds Token-only in Goal, while Token-only uses fewer target tokens.These variants change the ranking of eligible revisions; the reported checkpoint comparison is descriptive because training exposure and checkpoint selection differ.Source: §4.2 · pp. 7–8, Table 2 · p. 8
  • On AgentSense, TACT improves goal success and relationship scores over the initial student while using fewer target tokens and messages.The evaluation uses the same 500 instances from 100 templates.Source: Table 3 · p. 8, §4.2 · p. 8
  • TACT ranks highest among four methods in Goal across fixed Qwen and Llama partners and matched-policy self-play, on both All and Hard.With Llama it uses fewer target tokens than vanilla OPD; in every partner condition it uses more turns.Source: Table 4 · p. 9, §4.2 · p. 9
  • All three judges rank TACT first in Goal on All and Hard when scoring the same saved dialogues, although this does not establish agreement with human assessments.Kimi slightly favors vanilla OPD over TACT on Hard Avg.Source: Table 5 · p. 9, §4.2 · p. 9

5 CONCLUSION

Communication efficiency depends on both goal attainment and the communication needed to reach it. TACT trains from student actions, tests revisions through partner responses, and distills selected references; it improves goal outcomes while reducing communication on SOTOPIA and AgentSense.

5 CONCLUSION

  • TACT achieves the highest Goal among evaluated methods on SOTOPIA All and Hard with substantially fewer target tokens than SFT+SDPO, and improves AgentSense goal success with fewer tokens and messages.Component comparisons show higher Goal when expression and strategy supervision are combined, with distinct token-use and interaction-turn patterns.Source: §5 · p. 9
  • TACT starts from student-generated actions, proposes expression and strategy revisions, and samples partner replies before selecting a reference using local goal support and token cost.The selected reference guides on-policy distillation on the student’s original prefixes, so deployment requires only the student.Source: §5 · p. 9
  • The paper frames communication efficiency as reaching an agent’s goal while limiting the communication needed to reach it.Its supported conclusion is that agents should choose better actions and express them economically.Source: §5 · p. 9

ETHICS STATEMENT

Efficient goal pursuit is not sufficient evidence of socially appropriate behavior: private-goal proxies may miss harmful or unreliable conduct, and simulated results do not assure behavior with people.

ETHICS STATEMENT

  • Private-goal proxies can overlook coercion, omitted boundaries, or unreliable commitments.Goal pursuit alone should not be treated as evidence of socially appropriate behavior.Source: ETHICS STATEMENT · p. 10
  • Simulated results provide no assurance for deployment with people.Any human study requires its own consent and review procedure.Source: ETHICS STATEMENT · p. 10
  • Human studies require their own consent and review procedure.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

Reproduction and auditing rely on detailed run, training, and evaluation records. The reported tables distinguish several cohort and comparison types rather than treating them as interchangeable.

REPRODUCIBILITY STATEMENT

  • Reproduction requires role-visible prompts, exact model and tokenizer versions, raw tokens, scoring targets, branch seeds, selection records, probabilities, and checkpoints.The appendix specifies records needed to audit the training and evaluation pipeline.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Result tables distinguish recovered complete-panel endpoints, historical complete-case cohorts, within-trajectory comparisons, and development screening.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Exact run configurations and checkpoint bindings are retained with their evaluation records.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

The statement describes AI assistance, implementation settings, information access and supervised positions, and records retained for auditing. These details specify assistance and procedures without certifying social sufficiency.

AI USE STATEMENT

  • Generative AI assisted with research ideation and execution, code development, and manuscript organization, wording, and LaTeX editing.The authors remain responsible for verifying the method, implementation, citations, and evidence.Source: AI USE STATEMENT · p. 10

A.1 MODELS AND OPTIMIZATION

  • Table 6 summarizes verified school-framework run settings; training-generation and evaluation temperatures differ, and thinking generation is disabled.Exact model revisions, prompts, manifests, and checkpoint bindings remain in run records; shared settings do not remove exposure differences in Table 17.Source: Table 6 · p. 13, §A.1 · p. 13

A.2 INFORMATION ACCESS AND SUPERVISED POSITIONS

  • Specialists receive the target agent’s visible history and original action, while partner branches use partner-visible information without teacher identities or hypothetical future replies.Training loss covers the student’s generated action tokens and necessary endings, not history, partner tokens, fixed template tokens, or padding.Source: §A.2 · p. 13

A.3 AUDIT RECORDS

  • Audit records bind role-visible traces, tokens, probabilities, prompts, seeds, branches, targets, and selection decisions to training and evaluation checkpoints.These records support auditing but do not certify social sufficiency.Source: §A.3 · p. 13

A.4 EXACT SAMPLED-TOKEN ACTOR UPDATE

The sampled-token update distills teacher feedback on the student’s original generated tokens, with the selected reference appended only to the teacher context. The implementation also specifies token boundaries, batch handling, and a separate top-16 variant.

A.4 EXACT SAMPLED-TOKEN ACTOR UPDATE

  • At each original student token, the update uses teacher feedback conditioned on the selected reference and a clipped ratio comparing current and batch-frozen student probabilities.The selected reference is appended to the final user message; the scoring prompt does not separately append the original action or candidate branch reply.Source: §A.4 · pp. 13–14
  • The dynamic-batch actor computes token means within microbatches, then weights by action count; it clips accumulated gradient norm at 1.0 and takes one AdamW step.The implementation uses one epoch and one optimizer minibatch per collected action batch, with no entropy bonus or additional reference-policy KL loss.Source: §A.4 · p. 14
  • Scoring and updating share the same student actor and temperature, and cached feedback is tied to a frozen snapshot and invalidated after an update.The student, training partner, and goal scorer refresh between batches; teachers remain frozen and optimizer state is preserved.Source: §A.4 · p. 14
  • The separately implemented Student Top-16 variant weights teacher–student differences over the student’s 16 most probable tokens and is not the sampled-token update used in the main experiments.Its support-normalized weighting is neither exact full-vocabulary KL nor KL between renormalized 16-class distributions.Source: §A.4 · p. 14

A.5 ACTOR INFORMATION AND TOKEN BOUNDARIES

  • The complete serialized action and required EOS are supervised, while prompt tokens, other roles’ actions, fixed template positions, and padding are masked.Actions must contain exactly one JSON object with string fields action type and argument, using an environment-available action type.Source: §A.5 · p. 15
  • The OPD target is the student’s exact generated token sequence, scored with compatible teacher and student token IDs and templates rather than a retokenized proposal.The reference text is appended before rendering the teacher context; proposal-generation logits are not reused because they score different prefixes.Source: §A.5 · p. 15

B TRAINING AND SELECTION ALGORITHMS

This section describes TACT’s training and selection procedures, including specialist alternatives, partner-reply testing, reference distillation, and the prompts and scoring used. It also defines evaluation outcomes and communication costs, explains how missing outcomes and uncertainty are handled, and documents how to reproduce the runs and interpret comparisons.

B TRAINING AND SELECTION ALGORITHMS

  • TACT compares each original action with eligible specialist alternatives after one partner reply, selects by goal-support gain and token cost, then distills feedback on the student’s original tokens.The source interaction always advances with the original action; branch validation stops after one reply and does not establish final-outcome non-inferiority.Source: B TRAINING AND SELECTION ALGORITHMS · pp. 16–17

C.1 STUDENT AND PARTNER ACTION GENERATION

  • Student and partner actions use the same JSON action format and role-visible interaction history, with hidden profile fields masked by the environment renderer.The action-generation prompt requires an action instance with exactly the keys “action type” and “argument.”Source: §C.1 · p. 18

C.2 EXPRESSION AND STRATEGY SPECIALISTS

  • The expression specialist proposes a shorter action of the same type, while the strategy specialist may change intent or type to address the partner’s response and unresolved conditions.Only the expression action type is checked; other meaning-preservation requirements are prompt instructions, and neither specialist sees future dialogue or the partner’s hidden goal.Source: §C.2 · p. 19

C.3 GOAL-SUPPORT SCORING

  • Goal-support scores are model support for fixed goal text, not calibrated probabilities of achievement; the before and after calls differ only in visible interaction messages.The scorer averages target-token log probabilities, and no outcome sentence is sampled.Source: §C.3 · pp. 19–20

C.4 TEACHER-ONLY REFERENCE FOR DISTILLATION

  • For distillation, the teacher receives the selected action as an unexecuted, teacher-only reference, while the student’s original output prefix is teacher-forced.No candidate reply is added to dialogue history, and the student’s acting prompt is unchanged.Source: §C.4 · p. 20

C.5 PROMPTED OPD BASELINE

  • The prompted OPD baseline changes only teacher instructions, whereas the concise-prompt baseline adds an inference instruction without updating parameters.The OPD instruction emphasizes goal advancement and concise communication; the concise baseline asks the student to avoid unnecessary turns and tokens without compromising goal achievement.Source: §C.5 · p. 20, §C.6 · p. 20

D.1 DATASET CONSTRUCTION

  • Although variants share 1,129 configurations, the reported TACT checkpoint used 2,970 admitted nodes after 58 updates and 928 visits, unlike the other variants’ exposure.Expression-only and strategy-only each made 71 updates with 2,438 and 1,769 admitted nodes, respectively; checkpoint and exposure rules remain unequal.Source: §D.1 · p. 23

C.7 FINAL DIALOGUE JUDGE

  • Evaluation uses the target agent’s final dialogue-judge scores across 450 manifest cases, including all 70 Hard cases; action generators do not receive the judge’s full retrospective background.The judge evaluates seven dimensions, while target-agent results are taken from agent 1; the reported checkpoint tables use one complete 450-case evaluation.Source: §C.7 · p. 21, §D.1 · p. 23, §D.2 · p. 23

D.2 CHECKPOINT AND EVALUATION RUNS

  • The main-table checkpoint run recorded 446 first-pass ratings, one judge-retry rating, and three schema-clarification ratings, with candidate-selection statistics tied to its 2,970-node endpoint.Raw-output costs and episode-to-judge hashes were verified.Source: §D.2 · p. 23

D.3 SOCIAL SCORES AND COMMUNICATION COSTS

  • Avg is the unnormalized arithmetic mean of seven scores on the same case set, so it accompanies rather than replaces the individual dimension scores.All dimensions have higher-is-better scoring, and Avg does not define a social-sufficiency threshold.Source: Table 8 · p. 23, §D.3 · p. 24

D.4 PAIRING, MISSINGNESS, AND UNCERTAINTY

  • Evaluation binds scenario, roles, partner, judge, and seed; missing generation or judging remains distinct, and short low-scoring interactions are not counted as successes.Unscored endpoints are not assigned zero or retried until favorable; historical intervals use complete-case subsets, while complete-panel intervals and paired differences are computed separately.Source: §D.4 · pp. 24–25
  • Complete-panel recovery preserves valid outcomes, repairs serialization only under an exact-content check, and continues unrecoverable actions as none from the saved prefix.The final recovery added ratings across five methods while preserving the previous 2,192 valid ratings; this is post-hoc recovery, not uniform first-attempt evaluation.Source: §D.4 · p. 24

D.3 SOCIAL SCORES AND COMMUNICATION COSTS

  • Communication costs count raw generated action tokens and executed actions; Sotopia-RL uses a different tokenizer, so its token counts are model-native, not tokenizer-controlled.Tokens include serialized actions and ending tokens, while fixed prompts, judge calls, training, and discarded replay attempts are excluded; one turn is one executed action by either participant.Source: §D.3 · p. 24, §D.4 · p. 25

D.5 SOURCE VERSIONS AND REPRODUCTION ORDER

  • Reproduction binds model, tokenizer, templates, scenarios, decoding, and context settings before collecting and freezing each batch and saving update state.Evaluation also freezes the target checkpoint, partner, judge, and case manifest; configuration changes require separately attributable results.Source: §D.5 · p. 25

D.6 BASELINE TRAINING

The baseline training setups differ in data exposure and training partners, limiting direct loss-only comparisons. A prompted OPD baseline adds a teacher instruction without candidate proposals or outcome filtering, and the model-role setup distinguishes shared specialist parameters from separate scoring and distillation roles.

D.6 BASELINE TRAINING

  • Teacher SFT uses 27B target-role demonstrations with a fixed initial 4B partner, while vanilla OPD uses the same-batch student snapshot as partner; their different partners and admitted data prevent a loss-only controlled comparison.Teacher SFT uses 5,680 target actions and 355 updates; vanilla OPD uses 6,368 actions and 71 updates, and both traverse the 1,129-configuration schedule once.Source: §D.6 · pp. 25–26
  • The concise-prompt baseline adds a goal-oriented conciseness instruction without training, whereas SFT+SDPO applies segment-preference optimization after public SFT data.Sotopia-RL uses its published checkpoint without local retraining.Source: §D.6 · p. 26
  • Prompted OPD trains a fresh Qwen3.5-4B student with a frozen Qwen3.5-27B teacher whose prompt adds joint goal, strategy, and conciseness instructions, without proposals, IG gates, reference actions, or outcome filtering.The run completes 71 updates on 5,940 admitted student actions and is evaluated once on the original 450/70-case panel; it uses one training seed and one evaluation panel.Source: §D.6 · p. 26

D.7 MODEL ROLES

  • TACT specialists share a frozen backbone, while the IG scorer and distillation teacher are distinct roles.Source: Table 9 · p. 26

E AGENTSENSE TRANSFER EVALUATION

The AgentSense transfer evaluation samples 100 two-person templates and compares fixed checkpoints under an adapted dialogue protocol. Supplementary SOTOPIA analyses distinguish complete-panel uncertainty and historical development results from the selected main-table checkpoint.

E AGENTSENSE TRANSFER EVALUATION

F.1 FULL SOCIAL PERFORMANCE AND COMMUNICATION COST

  • The IG-only and Token-only comparisons train fresh students on the same 1,129 configurations, and evaluate only their final checkpoints against TACT’s selected checkpoint.Source: §F.1 · p. 28, Table 13 · p. 29

F.2 COMPLETE-PANEL UNCERTAINTY AND PAIRED DIFFERENCES

  • Complete-panel SOTOPIA intervals use scenario-cluster resampling, while paired comparisons resample identical scenarios and matched cases for both methods.Source: §F.2 · p. 28

F.3 DEVELOPMENT ENDPOINTS AND MAIN-TABLE REFERENCE

  • Development endpoints and the selected 10−5 checkpoint have different evaluation coverage and selection; the development results do not establish an untouched test set.Source: §F.3 · pp. 28–29
  • −0.405 (95% interval [−0.740, −0.067]; same context, n = 385) and −0.240 ([−0.589, 0.112]; reference context, n = 391) were historical late-minus-early Goal changes, so performance did not improve uniformly.Source: §F.3 · p. 29

F.4 TRAINING EXPOSURE

The training-exposure analysis distinguishes admitted nodes, dialogue configurations, and optimizer updates, while noting that comparable aggregate training costs have not been established.

F.4 TRAINING EXPOSURE

  • Node counts are outputs of collection and admission, not equal-compute budgets; comparable aggregate training costs have not been established.Source: §F.4 · pp. 29–30
  • Tables 14 and 15 provide complete-panel intervals and paired TACT-minus-comparator differences, with positive Goal and negative cost differences favoring TACT.Source: Table 14 · p. 30, Table 15 · p. 30
  • The 10−5 rows reproduce the complete 450/70-case evaluation, whereas development panels have different coverage and checkpoint selection.Source: Table 16 · p. 31
  • Training exposure treats a selected training node, a dialogue configuration, and an optimizer update as distinct units.Source: Table 17 · p. 31

F.5 CANDIDATE SELECTION AND DOWNSTREAM VALIDATION

TACT’s local candidate diagnostics show positive average downstream Goal differences for positive-IG-gain candidates, while cross-judge Goal rankings and partner comparisons reveal both consistent rankings and setting-dependent outcomes. The paper also marks qualitative case analysis as a needed but unreported empirical component.

F.5 CANDIDATE SELECTION AND DOWNSTREAM VALIDATION

  • Expression supplied 2,036 (68.6%) and strategy 934 (31.4%) of TACT’s 2,970 training references; ranking can decide only at 805 dual-eligible nodes.Source: Table 18 · p. 32, §F.5 · p. 32
  • On Random’s 841 dual-eligible nodes, 419 choices selected the strictly lower-IG/token candidate, including 168 Pareto-dominated choices that are not demonstrated downstream errors.Source: §F.5 · p. 33
  • +0.147 was the observed mean Goal improvement for positive-IG-gain candidates; the upper-quartile group improved by +0.244 versus +0.116 for other positive-gain candidates.Source: Table 20 · p. 33, §F.5 · p. 33
  • −0.005 ([−0.025, +0.016]) is the paired Goal difference between TACT and Random on random-cohort common nodes; teacher-gap candidate–student pairs show ρ = 0.256.Source: Table 21 · p. 34, §F.5 · p. 34, Table 22 · p. 34

F.6 PARTNER GENERALIZATION AND JUDGE SENSITIVITY

  • TACT ranks highest in Goal across all six partner–subset blocks, while its token and turn costs versus vanilla OPD vary by partner.Source: Table 23 · p. 35, §F.6 · p. 35
  • The partner panels are descriptive single-run comparisons and do not isolate partner effects from sampling variation or recovery procedures.Source: §F.6 · p. 36
  • All three judges rank TACT highest in Goal on All and Hard and agree on the All ordering, while absolute scores and lower-ranked methods vary by judge.Source: §F.6 · p. 36, Table 24 · p. 37

F.7 QUALITATIVE ANALYSIS

  • No empirical case analyses are asserted; any illustrative dialogues should not be treated as experimental observations.Source: §F.7 · p. 37
Loading 2609.35749v1…