PaperlayerView exact revision on arXiv

Source-linked AI summary

PsyEvo: A Personalized Counseling Agent That Self-Evolves at Test Time

Yuting Yan, Shihao Xu, Junhao Yu, Mingcong Zuo, Lu Chen, Nan Xiang, Haiyang Geng, Dongjie Tao, Minghao Wang

arXiv:2610.02885v1cs.AI

TL;DR

PsyEvo addresses the limits of offline training for adapting counseling to individual clients and ongoing therapeutic interaction. It combines client-specific skill selection, shared response-policy updates, and ordinal learning signals; in simulated-client PsychEval evaluation, it scores 7.684 Overall and exceeds every component variant across three matched runs.

  • Problem

    Offline model training alone offers limited room to adapt to individual clients or learn from ongoing therapeutic interaction, despite the need for scalable psychological support.

    Source: §1 · pp. 1–2

  • Method

    PsyEvo combines a per-client skill posterior, a cross-client shared response adapter, and SOCA-generated candidate preferences and trajectory credit over a frozen backbone.

    Source: §1 · p. 2

  • Results

    7.684 Overall on PsychEval; PsyEvo exceeds every component variant in each of three matched runs.

    Source: §3.3 · p. 6, §3.4 · p. 7

  • Takeaways & Limitations

    Matched PsychEval ablations support conditional component contributions, while the design lets individual intervention preferences evolve alongside a shared response policy with the backbone frozen.

    Source: §5 · p. 9

    Automatic evaluations use simulated clients and LLM judges, so their scores characterize model behavior rather than clinical efficacy; shared online cohort adaptation does not establish transfer to unseen clients.

    Source: §5 · p. 10

Abstract

from arXiv · show

Mental health disorders affect a substantial proportion of the global population, yet a persistent shortage of trained practitioners leaves the majority without adequate care. Large language model (LLM)-based counselors present a promising direction for delivering scalable conversational psychological support. Offline model training alone leaves limited room to adapt to individual clients or to learn from ongoing therapeutic interaction at test time. We introduce PsyEvo, an LLM-based counseling framework that enables both client-specific personalization and response-policy improvement at test time through three components: Hierarchical Bayesian Skill Policy (HBSP) personalizes what intervention to apply by maintaining a per-client skill posterior updated from session feedback; Inter-session Listwise Preference Optimization (LiPO) improves how the selected skill is expressed by updating a shared response adapter from cross-client preference evidence; and State-conditioned Ordinal Credit Assignment (SOCA) supplies candidate preferences and trajectory credit to the two components through consistency-checked comparisons and ordinal projection. In simulated-client evaluation with shared online cohort adaptation, PsyEvo obtains 7.684 Overall on PsychEval and exceeds every component variant in each of three matched runs. Removing individual components lowers mean overall score by 0.138--0.171 under the shared configuration, supporting conditional contributions within the complete scaffold. Our code is available at https://github.com/Lingxi-mental-health/PsyEvo

1 INTRODUCTION

LLM-based counseling offers a scalable conversational interface amid a substantial mental-health burden and counselor shortage. PsyEvo targets client-specific intervention selection and response-policy learning during test-time interaction.

1 INTRODUCTION

  • Nearly one billion people are affected by mental health disorders, while the available trained-counselor workforce falls far short of demand.LLM-based counselors are presented as a scalable conversational interface for psychological support.Source: §1 · p. 1
  • Existing LLM counseling approaches extend support across sessions, while differences in clients’ experiences, needs, and communication styles motivate adapting intervention selection and response policy.The paper frames these as complementary objectives: client personalization and cross-client counselor self-evolution.Source: §1 · pp. 1–2
  • PsyEvo combines a per-client skill posterior, a shared response adapter, and SOCA-generated preferences and trajectory credit over a frozen backbone.Its figure contrasts this test-time learning with static deployment and notes that shared-adapter updates require sufficient preference evidence.Source: §1 · p. 2, Figure 1 · p. 2

2 PSYEVO

PsyEvo separates client-private skill adaptation from population-shared response-policy learning while keeping the backbone frozen. HBSP and LiPO consume SOCA’s candidate preferences and trajectory credit to update those two learning paths.

2.1 PROBLEM FORMULATION

  • PsyEvo represents counseling as sequential decisions with a client-private memory and skill posterior alongside a population-shared adapter and evidence buffer.The backbone remains frozen, and HBSP, LiPO, and SOCA respectively personalize skill selection, improve skill expression, and construct learning signals.Source: §2.1 · pp. 2–3

2.2 HBSP

  • At each session’s end, HBSP updates the client’s skill posterior from SOCA-assigned trajectory returns, while posterior sampling preserves greater uncertainty for less-visited state–skill cells.The shared skill-prior table remains fixed during interaction; the posterior captures individual preferences.Source: §2.2 · p. 3

2.3 LIPO

  • LiPO updates a shared response adapter from SOCA-supported preference edges, weighting losses by edge support and normalizing within each candidate group.Accepted adapter versions are used only by sessions beginning after the update boundary, so response generation stays fixed within a session.Source: §2.3 · p. 4

2.4 SOCA

  • SOCA addresses compressed self-evaluation scores by accepting only pairwise directions supported in both presentation orders and discarding edges below a stability floor or creating cycles.The remaining edge strengths weight LiPO preferences; for HBSP, ordinal turn outcomes are projected and discounted across the realized trajectory to assign skill credit.Source: §2.4 · p. 4

3 EXPERIMENTS

Experiments evaluate PsyEvo on PsychEval and SummEval, using shared online cohort adaptation for counseling and human judgments for summary evaluation. PsyEvo scores 7.684 Overall on PsychEval, exceeds its component variants in matched runs, and receives the highest descriptive pilot rating among the counseling systems tested.

3.1 DATASETS AND BASELINES; 3.2 EVALUATION METRICS

  • The study evaluates a 100-client, five-modality PsychEval cohort, compares PsyEvo with eleven counseling and general-purpose baselines, and tests SOCA on SummEval against human judgments.PsychEval shares adapter updates across the online cohort, while each client’s skill posterior uses only that client’s feedback; Overall averages four metric families, and SummEval uses within-document Spearman correlations.Source: §3.1 · p. 5, §3.2 · p. 6

3.3 MAIN RESULTS

  • 7.684 Overall on PsychEval exceeds PsychAgent-32B by 0.336 and Gemini-3.8-Flash by 0.321; complete-system rankings are descriptive.PsyEvo’s shared online cohort-adaptation results average seeds 7, 17, and 27.Source: Table 1 · p. 6, §3.3 · p. 6
  • SOCA’s SummEval variant achieves a mean correlation of 0.508, surpassing PCA (0.489) by 0.019 and JDM (0.467) by 0.041.It has the highest Spearman correlation on coherence, fluency, and relevance, and ties for highest consistency correlation.Source: §3.3 · p. 6, Table 2 · p. 7

3.4 ABLATION STUDY

  • Removing LiPO lowers mean Overall by 0.171, removing SOCA by 0.152, replacing LiPO with pairwise DPO by 0.146, and removing HBSP by 0.138.PsyEvo exceeds every component variant in Overall in all three matched runs; these conditional contrasts do not decompose its full advantage over the base model.Source: Table 3 · p. 7, §3.4 · p. 7
  • Removing SOCA’s pairwise comparisons drops mean SummEval correlation by 0.041, with coherence most affected at −0.089.Removing pointwise scoring and repair evidence produces smaller decreases of 0.016 and 0.013, respectively.Source: Table 4 · p. 7, §3.4 · p. 7

3.5 HUMAN EVALUATION

  • In a descriptive pilot, three psychology experts rated simulated dialogues from 25 cases; PsyEvo’s mean Overall rating was 4.10, versus 3.58 for PsychAgent-32B.Ratings used a 1–5 scale; client improvement reflects impressions of simulated dialogue, not clinical change, and dimension-level agreement was low.Source: Table 5 · p. 8, §3.5 · p. 8

3.6 ANALYSIS

  • Across 100 clients and five modalities, HBSP rankings differ at shared states, with median within-modality disagreement ranging from 5.6% for CBT to 14.3% for PDT.Most slots agree; Figure 4’s trajectories cover changing sets of scored clients and do not isolate learning effects from other course dynamics.Source: §3.6 · p. 8, Figure 3 · p. 8, Figure 4 · p. 9

4 RELATED WORK

Prior work addresses counseling dialogue, longitudinal interaction, personalization, and feedback-based self-improvement, while PsyEvo combines client-private skill posteriors with a shared response adapter.

4 RELATED WORK

  • LLM counseling has progressed from within-session dialogue and report-grounded reconstruction to multi-session systems with memory and planning.PsyEvo combines client-private skill posteriors with a shared response adapter updated at test time.Source: §4 · p. 9
  • Personalization and test-time self-improvement have largely been explored separately, using user histories or parameter-efficient modules and feedback loops, respectively.LaMP and OPPU personalize generation; Self-Refine, Reflexion, and ExpeL support iterative self-improvement (Salemi et al., 2024; Tan et al., 2024; Shinn et al., 2023; Zhao et al., 2024).Source: §4 · p. 9
  • SOCA constructs ordinal evidence from consistency-checked comparisons without learning a reward model.PsyEvo’s LiPO objective uses SOCA support-strength weights and group normalization, unlike the gain and rank-discount weights of LiPO-Lambda.Source: §4 · p. 9

5 CONCLUSION AND LIMITATIONS

PsyEvo combines client-specific skill selection, cross-client response-policy improvement, and ordinal credit assignment with the backbone frozen. Its evaluation supports conditional component contributions, while remaining limited to simulated clients and shared-cohort adaptation.

5 CONCLUSION AND LIMITATIONS

  • PsyEvo combines HBSP for client-specific skill selection, LiPO for cross-client response-policy improvement, and SOCA for candidate preferences and ordinal trajectory credit.The backbone remains frozen while individual intervention preferences and the shared response policy evolve.Source: §5 · p. 9
  • Matched PsychEval ablations show consistent conditional component contributions across three runs, while SummEval separately examines evidence aggregation against human judgments.The ablations do not isolate the joint benefit of both learning paths or each SOCA mechanism.Source: §5 · pp. 9–10
  • Evaluations use simulated clients and LLM judges, so scores characterize model behavior rather than clinical efficacy.Shared online adaptation does not establish transfer to unseen clients.Source: §5 · p. 10
  • The human-rating pilot has limited agreement and remains descriptive; broader independent human evaluation is future work.Clinician-supervised settings also require assessment of crisis handling, privacy, and policy drift.Source: §5 · p. 10

A DATASET DETAILS

The appendix specifies the datasets and evaluation protocols used for counseling benchmarks and summary evaluation, including provenance and comparability qualifications.

A DATASET DETAILS

  • PsychEval evaluates a 100-client cohort across five therapy modalities, while MusPsy evaluates 100 test profiles over six sessions.PsychEval comprises 369 trajectories; MusPsy is a CBT-oriented synthetic corpus, and its reported supplementary-run configuration linkage remains unresolved.Source: A DATASET DETAILS · p. 13

B.2 SUMMARY-EVALUATION BASELINES

  • SummEval contains 1,600 summaries from 16 systems for 100 articles, and evaluator baselines are applied to these fixed outputs.The comparison covers six evaluator baselines; SOCA’s aggregation is assessed against expert and crowd-worker judgments.Source: A DATASET DETAILS · p. 13, §B.2 · p. 14

B.1 COUNSELING BASELINES

  • PsychEval comparisons include eleven baselines, but several scores are imported published references and retain their original evaluation conditions.The highest-mean comparison is descriptive, not a paired significance claim; matched within-scaffold interventions provide the component-contribution evidence.Source: §B.1 · pp. 13–14

C IMPLEMENTATION AND REPRODUCIBLE UPDATE SCHEDULE

The implementation evaluates learning within a shared online client cohort while preserving client-private states and session-fixed adapter versions. Supplementary details specify the learning-state updates, evidence handling, and reference implementation boundaries.

C IMPLEMENTATION AND REPRODUCIBLE UPDATE SCHEDULE

  • Within PsychEval runs, clients share one response adapter and preference buffer, while memories and skill posteriors remain client-specific.A session can use an adapter updated from earlier sessions of other cohort clients; the setup does not establish transfer to a disjoint cohort.Source: C IMPLEMENTATION AND REPRODUCIBLE… · p. 14
  • HBSP indexes skills with 54 slots formed from counseling stage, session progress, and keyword-based affect categories.Its shared skill table is fixed; client-specific residuals update under a Gaussian working model, which is an online approximation rather than an online-learned population hyperprior.Source: C IMPLEMENTATION AND REPRODUCIBLE… · p. 15
  • SOCA projects session-reward history to mean-preserving ordinal values and discounts accumulated trajectory credits.Realized-turn outcomes retrospectively assess emotional change, disclosure, collaboration, and concrete progress after a response.Source: C IMPLEMENTATION AND REPRODUCIBLE… · p. 15
  • Matched component variants retain the shared configuration but intervene on specified learning paths; the SOCA variant jointly changes candidate preferences and trajectory credit.The intervention does not isolate ordinal projection, and the fixed-adapter and no-HBSP variants each retain the other adaptation path.Source: C IMPLEMENTATION AND REPRODUCIBLE… · p. 16, Table 9 · p. 16

C.1 SUPPLEMENTARY IMPLEMENTATION DETAILS

  • The supplementary package implements HBSP, SOCA, and population LiPO operations separately from inference and benchmark evaluation.It is a reference implementation, not an export of the experimental training system; model-specific generation, feedback prompts, and physical adapter training come from a backend.Source: §C.1 · p. 16
  • SOCA converts two-order evaluator comparisons into weighted preference edges, while realized feedback supplies trajectory returns rather than candidate preferences.LiPO optimizes saved judgments; its group losses average retained-edge logistic losses before averaging across candidate groups.Source: §C.1 · p. 17
  • The MusPsy supplementary run uses CBT-only settings, but its correspondence to the reported MusPsy run remains unverified.The identified configuration checks update eligibility every two sessions and uses 16–160 preference pairs.Source: §C.1 · p. 18

D EVALUATION METRIC DETAILS

The evaluation uses distinct metric families for repeated counseling, MusPsy outcomes, and summary quality, with PsychEval aggregating four equally weighted score families. Its protocol also describes repeated-session coverage and the client-level aggregation used in the main tables.

D.2 MUSPSY

  • MusPsy evaluation reports WAI subscales and average, plus positive and negative PANAS means across six sessions for 100 clients.Higher WAI and positive PANAS and lower negative PANAS indicate better evaluator-assessed outcomes.Source: §D.2 · pp. 18–19

D.3 SUMMEVAL

  • SummEval measures macro-averaged Spearman correlation between automatic and human ratings for coherence, consistency, fluency, and relevance.Mean averages the four dimension-level correlations.Source: §D.3 · p. 19

D.1 PSYCHEVAL

  • PsychEval separates counselor-shared and counselor-specific intervention scores from client-shared and client-specific response scores, averaging the four families equally.It evaluates repeated interactions across CBT, BT, HET, PDT, and PMT.Source: §D.1 · p. 18, E PSYCHEVAL EVALUATION PROTOCOL · p. 19

E PSYCHEVAL EVALUATION PROTOCOL

  • PsychEval aggregation gives equal weight to clients after averaging each client's scored sessions, so longer counseling courses receive no greater weight.Session-wise means can reflect changing client composition over time.Source: E PSYCHEVAL EVALUATION PROTOCOL · p. 19

E.1 CROSS-RUN VARIABILITY

Across three matched seeds, PsyEvo ranked highest in Overall, while paired client-level intervals were positive for all four comparisons. The intervals describe variation conditional on the realized runs, not uncertainty over new learning schedules or cohorts.

E.1 CROSS-RUN VARIABILITY

  • 7.684 mean Overall (SD 0.028); PsyEvo scores highest in all three runs, and paired-difference 95% intervals lie above zero for all four comparisons.Table 10 reports across-seed means and sample standard deviations; Table 11 reports paired differences favoring PsyEvo.Source: §E.1 · pp. 19–20, Table 10 · p. 20, Table 11 · p. 20
  • Across seeds 7, 17, and 27, PsyEvo's paired run-level gains over DPO are 0.171, 0.146, and 0.121.The comparisons use matched runs on the same case set and average client-level scores equally.Source: §E.1 · p. 19
  • The client-level bootstrap intervals are conditional on three realized runs and do not capture uncertainty over new learning schedules or cohorts.Resampling does not rerun the coupled adaptation process, and three runs provide limited broader robustness evidence.Source: §E.1 · p. 19

F MECHANISM ANALYSIS OF THE THREE COMPONENTS

The mechanism analyses examine personalized skill rankings, adapter optimization records, and ordinal credit signals. The supplied evidence describes these mechanisms and traces, while noting where their correspondence to evaluation runs or causal effects is not established.

F.1 CLIENT-SPECIFIC SKILL SELECTION

  • HBSP's median skill-ranking disagreement across clients ranges from 5.6% to 14.3% across the five PsychEval modalities.The comparison uses 20 clients per modality at 54 shared state-space slots; tied slots are excluded from the disagreement denominator.Source: §F.1 · p. 20

F.2 PREFERENCE LEARNING AND ADAPTER CHANGES

  • Twelve accepted adapter updates over 2,080 preference pairs had finite gradients and changed the recorded adapter norm from 418.83 to 421.14.The trace establishes optimization and parameter change, but its correspondence to main evaluation runs and pair accounting unit are unspecified; quality is assessed by subsequent session outcomes, not training-loss reduction.Source: §F.2 · p. 20
  • DPO can learn from a valid preference direction without a large judge-score gap, while SOCA weights only retained acyclic-graph edges by bidirectional-consistency strengths.Score compression affects fixed-threshold pair selection or losses weighted by score gap.Source: §F.2 · p. 20

F.3 ORDINAL SIGNAL AND DELAYED CREDIT

  • Internal rewards span 0.569–0.779 across 63 sessions, and ordinal-credit standard deviations span 0.104–0.127 across twelve updates.These trace diagnostics indicate a non-degenerate internal signal; the corresponding distributions are shown in Figure 5.Source: §F.3 · p. 20, Figure 5 · p. 21
  • Discounted propagation can assign an early skill choice credit associated with later progress, while reducing the contribution of more distant outcomes.Isolating the effect would require a controlled intervention on the propagation rule.Source: §F.3 · p. 21

G REWARD SEPARATION AND JUDGE AGREEMENT

The evaluation separates internal learning feedback from external transcript evaluation and reports agreement between automated judges and human raters. The results include positive judge rank correlations, while the human evaluation is limited in scope and has dimension-level agreement constraints.

G REWARD SEPARATION AND JUDGE AGREEMENT

  • External scores evaluate frozen transcripts, while internal feedback supplies candidate preferences and trajectory returns; the write-path audit checks whether external scores enter training inputs or updates.Table 12 identifies feedback-channel recording units and the modules that consume them; only the first two channels enter learning.Source: G REWARD SEPARATION AND JUDGE AGREEMENT · p. 21, Table 12 · p. 21
  • Qwen3-32B and Gemini-3-Flash show positive rank correlations across four dimensions, ranging from 0.480 to 0.632 with confidence intervals above zero.The agreement analysis covers 1,468 matched sessions; positive correlations do not establish equivalent calibration, independent feedback, or clinical validity.Source: G REWARD SEPARATION AND JUDGE AGREEMENT · p. 21, Table 13 · p. 22
  • On 75 sessions from 40 clients, client-grounded reward correlates more strongly with external Gemini C-Shared than the local supervisor checklist.Pearson and Spearman correlations are 0.564 and 0.477 for client-grounded reward, versus 0.178 and 0.157 for the checklist.Source: G REWARD SEPARATION AND JUDGE AGREEMENT · p. 22

H.1 HUMAN EVALUATION PROTOCOL

  • Three experts independently rated 75 sessions from 25 matched cases, with system identities anonymized and presentation positions randomized.Client improvement was rated once per case after experts read three sessions from that case.Source: §H.1 · p. 22, §H.2 · p. 22

H.3 AGGREGATION

  • Human overall ratings equally weight six dimension means: the first five aggregate expert session ratings, while client-improvement ratings aggregate expert case ratings; results are reported to two decimal places.Source: §H.3 · p. 22

H.4 INTER-RATER AGREEMENT

  • Dimension-level human-rater agreement is limited, and the composite reliability interval is not a paired system-difference interval.The human comparison also does not include the strongest general-purpose systems from the automatic evaluation.Source: §H.4 · p. 23, Table 15 · p. 23

H.5 HET-860: CLIENT-SPECIFIC ADAPTATION ACROSS THIRTEEN SESSIONS

The HET-860 case shows client feedback changing intervention priorities selectively across recurring states, alongside progress over thirteen sessions. The CBT-412 case traces retained workplace context and evolving interventions across five sessions, while the failure cases flag risks from tied scores and erroneous memory.

H.5 HET-860: CLIENT-SPECIFIC ADAPTATION ACROSS THIRTEEN SESSIONS

  • The top-ranked skill set changed in 16 of 32 visited states and remained unchanged in the other 16, showing state-selective adaptation.Figure 6 compares skill rankings before Session 1 and after Session 13 across those states.Source: §H.5 · p. 23, Figure 6 · p. 24
  • For core intervention, middle-session progress, and anxiety, sixteen updates raised HET1's client-specific posterior mean from zero to 4.404 and reversed its ranking against HET3.The HET1-minus-HET3 value difference shifted from −0.430 to +3.974 while shared values stayed unchanged; HET1 was selected in Session 10.Source: §H.5 · p. 24
  • Across 384 skill decisions and 384 posterior updates, 282 decisions read a nonzero client-specific residual, and adding the posterior mean changed the highest-valued skill at 66 decision points.Updated state–skill entries grew from 17 after Session 1 to 72 after Session 13.Source: §H.5 · p. 24, Table 16 · p. 25
  • External WAI rose from 6.250 to 7.500, and mean PANAS and SRS rose from 5.188 to 6.875 over the trajectory.The client-shared score peaked at 7.562 in Session 9, then declined but ended above its initial value.Source: §H.5 · p. 24, Table 16 · p. 25

H.6 A CLIENT-SPECIFIC COURSE: CBT-412

  • In CBT-412, the recorded skills progressed from emotion regulation to trigger monitoring, cognitive distancing, and self-supportive language, while the client began remaining in triggering situations despite residual discomfort.Retained workplace triggers, bodily cues, and beliefs connected the five-session course to memory and both learning paths.Source: §H.6 · pp. 24–25

H.7 FAILURE CASES AND MEMORY DEPENDENCE

  • Fully tied score groups have no recoverable strict order, and erroneous cross-session memories can reinforce mistaken goals or beliefs in later skill selection and candidate evaluation.The authors recommend recording variance and skip decisions for ties, and preserving the source of client corrections in retained profiles.Source: §H.7 · p. 25

I COMPUTE, DEPLOYMENT, AND SAFETY

The appendix describes online training overhead and deployment alternatives, states limits on what ordinal properties guarantee, and qualifies supplementary comparisons. It also tests a standardized fusion of pointwise, pairwise, and repair evidence against human summary rankings.

I COMPUTE, DEPLOYMENT, AND SAFETY

  • The measured training trace took 2,217.6 seconds over twelve rounds, or 36.96 minutes, excluding dialogue serving and shadow generation.The mean trainer time was 3.08 minutes per round on four trainer GPUs; it does not establish response latency or end-to-end session time.Source: I COMPUTE, DEPLOYMENT, AND SAFETY · p. 26
  • Synchronous adapter activation makes update timing auditable but can delay the next session; asynchronous and uncertainty-triggered alternatives require separate latency and policy-version measurement.Clinical efficacy and safety also require prospective evaluation with professional oversight, privacy protections, and crisis-handling assessment.Source: I COMPUTE, DEPLOYMENT, AND SAFETY · p. 26

J UPDATE SCHEDULE AND ORDINAL PROPERTIES

  • Tie-aware ranks and tiers for realized trajectories remain invariant under strictly increasing transformations of turn scores, but projected distances need not remain invariant.Quantization can introduce ties, and a non-monotone evaluator can reverse an ordering; SOCA candidate edges instead use a two-order consistency rule.Source: J UPDATE SCHEDULE AND ORDINAL PROPERTIES · pp. 26–27

K.1 SUPPLEMENTARY MUSPSY COMPARISON

  • Supplementary MusPsy figures are not evidence for the core method comparison because the main PsyEvo run is not linked to that configuration and some baselines lack complete row-level sourcing.The authors retain the numerical comparison without claiming superiority or robustness.Source: §K.1 · p. 27, Table 19 · p. 27

K.2 MULTI-EVIDENCE AGGREGATION BEYOND COUNSELING

  • For SummEval, SOCA ranks summaries by combining standardized pointwise scores, pairwise win rates, and accepted-repair signals, with fixed αd = βd = 1 across dimensions.Only accepted repairs contribute to the repair channel; an accepted improvement counts as evidence of a defect in the original summary.Source: §K.2 · pp. 27–28
  • SummEval measures agreement with human rankings, but the reported mean correlation differences are descriptive because paired document-level uncertainty is not reported.The counseling SOCA ablation tests the combined contribution of consistency-checked candidate preferences and ordinal trajectory credit.Source: §K.2 · p. 28

Related papers

Loading 2610.02885v1…