acceptodds
Under review as a conference paper at ICLR 2027

NeedTracer: Tracking Dynamic User Needs with Focus Anchors in Multi-Turn Emotional Support

Abstract

In multi-turn emotional support, accurately understanding user needs is key to effective emotional support. Existing methods can continuously update implicit needs through ongoing interaction, but they have not explicitly determined whether observed need changes reflect further confirmation and refinement of the current need or indicate that a new need stage has formed. Consequently, when a cross-stage need transition occurs, the model may still organize subsequent support based on the previous need and gradually deviate from the user's current need. To address this issue, we propose NeedTracer, a dynamic user need tracking framework for multi-turn emotional support that aims to continuously identify the user's current need and its stage-wise changes. NeedTracer identifies and maintains the focus anchor based on user-side evidence accumulated through ongoing interaction, representing the specific unresolved concern that the user currently prioritizes, and uses it to organize within-stage updates and cross-stage transitions of the current user need. For within-stage updates, when the focus anchor remains stable, NeedTracer updates the current user need based on continuously arriving user-side evidence to capture further confirmation and refinement of the need within the same need stage. For cross-stage transitions, after the current need has received stage-wise sufficient support, NeedTracer confirms a new focus anchor based on new event developments and changes in the user's concern, and accordingly enters a new need stage. For support generation, NeedTracer uses a Theory of Mind (ToM)-based evaluator to select the support response that best matches the current user state and current user need. Experimental results show that NeedTracer consistently improves performance across multiple models, benchmarks, and long-horizon interaction settings, validating focus-anchor-based dynamic user need tracking.

Primary area:
foundation or frontier models, including LLMs
Keywords:
User Need Tracking, Large Language Models, Multi-Turn Interaction, Emotional Support Dialogue

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GzVaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

NeedTracer tracks dynamic user needs in multi-turn emotional support by distinguishing within-stage need updates (Keep/Deepen/Correct) from cross-stage transitions, which require sufficient need satisfaction and a consistently supported new focus anchor over two consecutive turns. It maintains global and current user states, generates preference-constrained and free-form responses, and selects them via a Theory-of-Mind-based evaluator. Experiments show gains in support effectiveness, interaction efficiency, and emotional trajectory metrics, supported by human evaluations and ablations.

Critical strengths
  1. The paper tackles a real challenge in emotional support: telling a need update apart from a genuine switch to a new need.

  2. NeedTracer offers a structured way to track changing user needs and adjust support responses accordingly.

  3. The paper tests multiple LLM backbones and benchmarks, and ablation studies plus human evaluations back up the method's effectiveness.

Critical weaknesses
  1. The paper doesn't show clearly what the focus anchor adds on its own. The gains could come from extra reasoning steps and longer prompts. A compute-matched comparison without focus anchors would help.

  2. Many key decisions depend on LLM judges, which are subjective. We need more on cross-model agreement, false transitions, and missed transitions.

  3. Waiting two turns may slow down real transitions. The system may keep addressing an old concern. More analysis of transition delay and different thresholds would help.

  4. Better responses don't prove the need tracking is accurate. A dedicated test with annotated stages and transitions, reporting precision, recall, F1, and detection delay, is needed.

  5. The compute cost is not well reported. Fewer turns doesn't mean cheaper. LLM calls, tokens, latency, and cost vs. baselines should be given.

  6. Results may not generalize. Simulated users and LLM evaluators can bias results, and three seeds don't show significance. Real-user or independent data, paired significance tests, and evaluator-bias analysis would help.

Questions for authors

N/A

2 comments

We thank the reviewer. NeedTracer anchors the highest-priority unresolved concern to distinguish Need Identification Uncertainty from Genuine Need Transition caused by satisfaction or priority change, stabilizing current user need tracking.

Q1: What does focus anchor add beyond extra reasoning?

A1: Independent contribution. Existing emotional support datasets lack comprehensive annotations for evaluating NeedTracer's ability to distinguish Need Identification Uncertainty from Genuine Need Transition and continuously track the current user need through a focus anchor in multi-turn interactions. Therefore, we select 300 dialogues (8–30 turns) from ESConv, ETrajEval, and ExTES, and annotate ground truth (GT) and supporting evidence for focus anchors, current user needs, and need-stage transitions to evaluate the independent contribution of the focus anchor and dynamic need tracking (Table R1).

w/o focus anchor retains stages, current/global states, Keep/Deepen/Correct, and two-turn confirmation, removing anchor/priority constraints. The base model, candidates, ToM reranking, memory, and compute ceiling are fixed.

Table R1. 300 dialogues; mean ± std over 10 seeds (Need Acc./Sat. in %).

MethodNeed Acc.↑Need Sat.↑Trans.↓RT(s)↓Calls↓Tokens/turn↓
Baseline35.7±0.753.5±1.311.9±0.120.9±0.193.7±0.07127±15
w/o focus anchor66.9±0.661.8±0.87.8±0.431.0±0.314.5±0.11165±26
NeedTracer (Full)78.6±0.382.6±1.05.6±0.391.2±0.235.6±0.15193±18

Versus w/o anchor: +11.7% Need Acc., +20.8% Need Sat. (absolute); 28.2% fewer transitions.

Q2: How reliable are judges and transition detection?

A2: Human agreement and errors. On 150 cases, Claude Opus 5.5, Gemini 4 Argon, and DeepSeek-V4-Pro are compared with three-annotator majority labels.

Table R2. LLM–human agreement (%); 150 cases.

DimensionOriginalUpdated
User Need81.689.1
focus anchor83.486.3
Stage-wise Need Satisfaction79.583.4
Cross-stage Need Transition85.992.3
Macro Average82.687.8

For transition testing, 200 independent dialogues (8–20 turns) cover 100 Satisfaction-driven (S) and 100 Priority-driven Override (O) cases, with three-annotator labels and expert adjudication.

Table R3. Transition P/R/F1 (%); 200 separate dialogues.

MethodS-PS-RS-F1O-PO-RO-F1
Baseline45.146.845.951.553.452.4
UKA68.975.772.173.672.172.8
w/o focus anchor72.573.272.869.173.571.2
w/o Override76.079.577.772.475.874.1
NeedTracer (Full)81.783.482.580.382.981.6

False-discovery (1−P)/miss (1−R) rates: 18.3%/16.6% (S) and 19.7%/17.1% (O).

Q3: Does two-turn confirmation delay transitions?

A3: Stable confirmation and override. Studies support feedback-based belief updates (Bohus & Rudnicky, 2005), multi-turn goal tracking (Williams et al., 2013), and intent clarification (den Hengst et al., 2024). Two consecutive user-feedback turns filter isolated mentions and within-concern Keep/Deepen/Correct changes. Satisfaction-driven Transition follows sufficient support and a new concern; Priority-driven Override updates the anchor when evidence and feedback confirm a higher-priority concern despite unmet needs. Unmet needs and anchors remain in long-term memory for future re-confirmation if priority returns. Revised Algorithm 1 adds Override; O-F1: 74.1%→81.6% (Δ=7.5%). Future work examines W=1/2/3, detection delay, and interrupted-need recovery.

Q4: Is need-stage tracking independently evaluated?

A4: Dedicated transition test. A need stage spans a focus anchor’s valid interval; Keep/Deepen/Correct updates remain within it. On R3’s 200 independent dialogues, cross-stage F1 reaches 82.5% (S)/81.6% (O), +9.7%/+10.4% over w/o focus anchor, independent of response quality.

Q5: What is the compute cost?

A5: Measured overhead. On R1’s 300 dialogues, Full adds 0.2 s RT, 1.1 calls, 28 tokens/turn versus w/o anchor. The +11.7% Need Acc./+20.8% Need Sat. gains justify this modest overhead for better need tracking.

Q6: How robust are findings across data, judges, and seeds?

A6: Multi-source validation. R1: 300 ESConv/ETrajEval/ExTES dialogues, 10 seeds; R2: 150 judge–human agreement cases, three model families; R3: 200 independent human-adjudicated transition dialogues. Across seeds: 78.6%±0.3% Need Acc., 82.6%±1.0% Need Sat., 5.6±0.39 transitions. Future work extends real-user and long-dialogue evaluation, paired tests, and evaluator sensitivity.

References

[1] Bohus, D., & Rudnicky, A. I. (2005). Constructing Accurate Beliefs in Spoken Dialog Systems. IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU 2005). Paper.

[2] Williams, J., Raux, A., Ramachandran, D., & Black, A. (2013). The Dialog State Tracking Challenge. Proceedings of SIGDIAL 2013, 404–413. Paper.

[3] den Hengst, F., Wolter, R., Altmeyer, P., & Kaygan, A. (2024). Conformal Intent Classification and Clarification for Fast and Accurate Intent Recognition. Findings of ACL: NAACL 2024, 2412–2432. Paper.

Reviewer_7UCqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NeedTracer, a framework for tracking dynamic user needs in multi-turn emotional support. It introduces focus anchors to distinguish within-stage need refinement from cross-stage need transitions. Experiments across four LLM backbones and multiple benchmarks demonstrate consistent improvements over existing methods.

Critical strengths

(1) The paper addresses the important problem of distinguishing within-stage need refinement from cross-stage need transitions. The focus-anchor mechanism provides an explicit framework for tracking evolving user needs in multi-turn emotional support.

(2) NeedTracer demonstrates strong performance across multiple LLM backbones and benchmarks (Table 1). The human evaluation (Tables 2-3) and ablation studies (Table 5) further support the effectiveness of the proposed framework.

Critical weaknesses

(1) Unclear effectiveness of the anchor-transition mechanism, particularly in short dialogues

The paper proposes switching to a new focus anchor only after the current need has been sufficiently supported (Sec. 3.4), but there is little empirical evidence on how often such transitions actually occur, especially in benchmark dialogues capped at 8-10 turns, or on whether the method can handle priority shifts that occur while a previous need remains unresolved. Since there is no analysis of performance by the number of transitions and the ablation (Table 5) is limited to ETrajEval, it is difficult to determine whether the performance gains stem from the transition mechanism.

(2) Computational cost and LLM-call overhead are not measured

The proposed framework involves multiple LLM calls per turn for candidate generation, internal evaluation, and verification. However, the resulting inference overhead, such as latency, token consumption, and API cost, is not reported, and efficiency is measured only by the number of dialogue turns. It is therefore not established whether the performance gains over simple prompting or direct generation justify the additional computational cost, which makes the practical value of the method difficult to assess.

Questions for authors

(1) Focus-anchor transitions (Sec. 3.4, Algorithm 1, Tables 1 and 6). Given the consecutive confirmation requirements and the 8-10 turn limits of the benchmarks, how frequently do focus-anchor transitions occur? How does performance vary with the number of transitions?

(2) Transition conditions (Sec. 3.4, Table 3, Appendix E). How does NeedTracer handle cases where the user's primary concern changes before the previous need reaches stage-wise satisfaction? Are such cases covered in the evaluation?

(3) Ablation settings (Table 5, Algorithm 1, Appendix A.3.5). Could the authors clarify how the ablations modify Algorithm 1, particularly the difference between History-based Need Tracking and w/o focus anchor?

(4) Inference overhead (Secs. 3.3–3.5, Table 1). Could the authors report inference latency, token consumption, and API costs relative to the baselines, beyond the dialogue-turn efficiency reported in Table 1?

(5) Necessity of candidate-based evaluation (Secs. 3.3–3.5, Table 3, Appendix E). How does direct LLM prediction without candidate generation and evaluation affect downstream performance, given that Appendix E reports only agreement-based sensitivity results?

2 comments
AuthorsOpenReview ↗

We thank the reviewer for the constructive feedback. We address the key concerns through focus anchor transition analysis, Priority-driven Override validation, controlled ablations, computational overhead evaluation, and candidate-based inference comparisons. Experiments on extended dialogues and independent transition cases demonstrate the effectiveness of focus anchor in improving need identification, reducing unnecessary transitions, and handling genuine priority shifts.

We thank the reviewer for the constructive feedback. Through a focus anchor, NeedTracer distinguishes Need Identification Uncertainty (insufficient evidence about the current user need) from Genuine Need Transition (need satisfaction or a change in concern priority). This prevents repeated need identification that weakens need satisfaction. New evidence supports our original findings.

W1 & Q1–Q2: Focus-anchor value and transitions. A1: The focus anchor represents the highest-priority unresolved concern. Keep/Deepen/Correct updates the current user need within this boundary. A transition occurs after sufficient support and a new concern, or when another concern becomes higher priority. Unlike direct need selection, the anchor distinguishes evolving need understanding from an actual stage change.

(1) Satisfaction-driven Transition: Update anchor after sufficient support and a shift to another unresolved concern. (2) Priority-driven Override: Update it if sufficient evidence and user feedback confirm a new top-priority unresolved concern, despite unmet prior needs. Retain the unmet need and its focus anchor in long-term memory for rechecking with user feedback if reprioritized.

Existing datasets lack gold labels for current psychological needs, focus anchors, and need-stage transitions. We therefore annotate ESConv/ETrajEval/ExTES dialogues and evaluate 300 dialogues (8–30 turns) under shared settings (10 independent seeds; mean ± std).

Table R1. Performance, transitions/dialogue, cost/response (10 seeds).

MethodNeed Acc.↑Need Sat.↑Trans./dlg.↓RT(s)↓Calls↓Tokens/turn↓
Baseline35.7%±0.7%53.5%±1.3%11.9±0.120.9±0.193.7±0.07127±15
w/o focus anchor66.9%±0.6%61.8%±0.8%7.8±0.431.0±0.314.5±0.11165±26
NeedTracer (Full)78.6%±0.3%82.6%±1.0%5.6±0.391.2±0.235.6±0.15193±18

Full improves Need Acc./Need Sat. by 11.7/20.8 percentage points and reduces transitions by 28.2% versus w/o focus anchor. Original Table 5 reports BEL 70.9±1.3 (Full), 59.1±2.1 (w/o cross-stage transition), 57.2±1.7 (w/o focus anchor).

For direct transition evaluation, we annotate 200 independent dialogues (8–20 turns): 100 Satisfaction-driven (S) and 100 Priority-driven Override (O), including shifts with unmet previous needs. Three annotators label anchors, stage spans, needs, and transition evidence; experts adjudicate.

Table R2. Transition P/R/F1 (%) and response RT (s).

MethodS-PS-RS-F1S-RTO-PO-RO-F1O-RT
Baseline45.146.845.90.951.553.452.41.2
UKA68.975.772.12.273.672.172.81.8
w/o focus anchor72.573.272.81.069.173.571.21.5
w/o Override76.079.577.71.472.475.874.11.3
NeedTracer (Full)81.783.482.51.280.382.981.61.2

Full achieves 82.5%/81.6% F1 on S/O; O-F1 exceeds w/o Override by 7.5 percentage points. These results support priority shifts with unmet previous needs and corroborate the original mechanism.

W2 & Q4: Computational overhead. A2: Table R1 reports per-response latency, LLM calls, and tokens/turn. Compared with w/o focus anchor, NeedTracer adds 0.2 s, 1.1 LLM calls, and 28 tokens/turn, while improving Need Acc. and Need Sat. by 11.7 and 20.8 percentage points and reducing average need transitions by 28.2%. These gains demonstrate the value of focus-anchor-guided tracking relative to its modest computational overhead. Further efficiency optimization remains future work.

Q3: Algorithm 1 ablations. A3: Full uses Keep/Deepen/Correct, sufficiency after two consecutive Keep confirmations, both transition paths, and current/global states. Static Need Inference predicts only initially; History-based Need Tracking re-infers each turn without anchors, stages, or stage states; w/o within-stage need update disables Deepen/Correct; w/o stage-wise need satisfaction removes the satisfaction gate but retains Override; w/o cross-stage need transition disables both paths; w/o focus anchor removes anchor/priority assessment but keeps stages, two-level states, within-stage updates, and sufficiency checks, using need changes plus two confirmations for stage transitions. Thus, History-based Need Tracking removes stage management; w/o focus anchor retains it. Response candidates, ToM reranking, and memory remain fixed.

Q5: Candidate-based inference. A4: Direct Prediction removes candidate generation and comparative evaluation; candidate-based inference uses K=3.

Table R3. LLM–human agreement (%).

DimensionDirect PredictionCandidate-based (K=3)
User Need63.681.6
Focus Anchor69.583.4

Candidate comparison improves User Need and Focus Anchor agreement by 18.0 and 13.9 percentage points, respectively, supporting more reliable judgments.

Reviewer_7UCqOpenReview ↗

Response to Author Rebuttal

Thank you for the additional experiments and clarifications. The response provides some useful details on the ablation settings and computational costs, but my main concerns remain.

The Priority-driven Override described in the rebuttal is not reflected in the original Sec. 3.4 or Algorithm 1, making it unclear whether the reported results correspond to the original method. Additionally, Table R1 aggregates dialogues of 8–30 turns and does not clarify how frequently anchor transitions occur or contribute to performance in the shorter ESConv and ExTES settings.

For these reasons, my assessment remains unchanged.

Reviewer_g5vwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes NeedTracer, a framework for tracking dynamic user needs in multi-turn emotional support. Its central idea is to distinguish refinement of an existing need from transition to a new need by maintaining a “focus anchor,” representing the user’s currently prioritized unresolved concern. The framework combines global and stage-specific user states, candidate-based need inference, satisfaction-gated anchor transitions, and ToM-based response selection. Evaluation covers four LLM backbones, four benchmark settings, three independent runs, and human assessments by three annotators. The reported results show improvements in support success, interaction length, human-rated dialogue quality, and long-horizon emotional outcomes.

Critical strengths
  1. A meaningful formulation of dynamic need tracking. Distinguishing within-concern refinement from changes in the user’s prioritized concern addresses an important problem in sustained emotional support. The explicit focus-anchor representation makes these decisions inspectable and provides a useful basis for analyzing conversational adaptation.

  2. Broad and consistently positive reported results. The evaluation spans four backbones and multiple interaction settings, uses separate simulation and external evaluation components, and includes blinded human assessment. The improvements are substantial in several comparisons, supporting further investigation of the approach.

Critical weaknesses
  1. The term "Stage" lacks sufficient justification. Are they explicitly tagged by the original dataset? Or are they implicitly implied by the approach methodlogy? This is important because the key idea of this paper contains both "cross-stage" and "within-stage", but there is no stage-specific quantitative analysis or case analysis.
  2. The method is guided by the LLM-based evaluator. However, the foundation of evaluators are equivalent (Deepseek-Flash) or weaker (GPT-5.6-luna VS GPT-5.6-Sol). This worries me since the weaker judger may provide misleading results; and the same series of player and judger may also provide bias evaluations.
  3. It seems the methdology is training-free and mainly prompt-based. However, the alogrithm is complicated (Algorithm 1). I find no efficient analysis or RT record which justify that this method can implement a practical dialogue system (instead of forcing the users to wait a long time before the response).
  4. It is difficult to figure out that how the case in Figure (left) spports the two needs below. It seems that the highlighted assistant response and the subsequent user queries cover need 1 only.
  5. The long-horizon metric is mathematically inconsistent with its stated interpretation. Appendix A.2 presents Eqs. (12) and (13) as equivalent, but a trajectory (0->1->0) yields ETV (=0.5) under Eq. (13), whereas Eq. (12) yields zero because the upward and downward transition probabilities are equal. Thus, Eq. (13) can reward oscillation without net emotional recovery, undermining the interpretation of Tables 4–5 as evidence of more stable emotional trajectories or greater resistance to regression.
  6. The central transition mechanism lacks sufficiently independent, targeted validation. Algorithm 1 permits an anchor transition only after the previous need is confirmed and sufficiently satisfied, although users may reprioritize concerns before the previous need is resolved. Table 3 reports aggregate agreement rather than transition precision, recall, and detection delay; moreover, Appendix E selects default hyperparameters using the same 150 test cases used for evaluator-reliability assessment. Consequently, the reported agreement does not establish robust generalization to genuine need transitions, particularly those that violate the framework’s sequential-satisfaction assumption.
  7. Some typos. For examples, the last three lines of Table 7 have bad symbols.
Questions for authors
  1. Which ETV definition was actually used to produce Tables 4–5? Please provide the implemented calculation and a worked trajectory example, explain the discrepancy between Eqs. (12) and (13), and clarify whether correcting it changes the reported rankings.
  2. Were (K_N=3), (K_A=3), and (W=2) selected using the same 150 cases on which the final agreement scores are reported? If so, are independent held-out estimates available, and could annotators reject all candidates when none adequately represented the user’s need?
  3. What does the implemented system do when the user clearly prioritizes a new concern while the previous need remains unsatisfied? Is there an override to Algorithm 1’s satisfaction gate, and do the existing evaluation cases provide evidence about transition accuracy and delay in this situation?
  4. Ablation comparability: Do History-based Need Tracking and the Table 5 ablations retain the same response-candidate generation, ToM reranking, memory capacity, and inference budget as the full system? Clarifying these controls is necessary to attribute the improvements specifically to focus-anchor-based tracking.
1 comment

We thank the reviewer for the constructive feedback. Through a focus anchor, NeedTracer distinguishes Need Identification Uncertainty (insufficient evidence about the current user need) from Genuine Need Transition (need satisfaction or changed concern priority), avoiding repeated need identification that weakens need satisfaction.

W1: Need-stage definition and validation. A1: A need stage shares one focus anchor (highest-priority unresolved concern). Keep/Deepen/Correct updates the current need within the stage; an anchor change marks cross-stage transition. Existing datasets lack gold need, anchor, and stage-transition labels. We thus annotate ESConv/ETrajEval/ExTES dialogues to evaluate these decisions.

W2: Evaluator capability and bias. A2: We compare Claude Opus 5.5, Gemini 4 Argon, and DeepSeek-V4-Pro against three-annotator majority labels (150 cases).

Table R1. LLM–human agreement (%): original vs. updated judges.

DimensionOriginalUpdated
User Need81.689.1
Focus Anchor83.486.3
Stage-wise Need Sat.79.583.4
Cross-stage Transition85.992.3
Macro Avg.82.687.8

Macro agreement rises 82.6%→87.8%; original blind human evaluation (Table 2) adds independent evidence.

W3 & Q4: Efficiency and comparable ablations. A3: We evaluate 300 annotated dialogues (8–30 turns) across the three datasets under shared settings (10 seeds; mean ± std).

Table R2. Performance, transitions/dialogue and cost/response (10 seeds).

MethodNeed Acc.↑Need Sat.↑Trans./dlg.↓RT(s)↓Calls↓Tokens/turn↓
Baseline35.7%±0.7%53.5%±1.3%11.9±0.120.9±0.193.7±0.07127±15
w/o focus anchor66.9%±0.6%61.8%±0.8%7.8±0.431.0±0.314.5±0.11165±26
NeedTracer (Full)78.6%±0.3%82.6%±1.0%5.6±0.391.2±0.235.6±0.15193±18

Full improves Need Acc./Need Sat. by 11.7/20.8 percentage points and 28.2% fewer transitions over w/o focus anchor, for +0.2 s, +1.1 calls, +28 tokens/turn: valuable gains relative to the cost. Ablations share the model, response candidates, ToM reranking, memory, and inference budget. History-based Need Tracking lacks stages; w/o focus anchor retains stage management and within-stage updates but removes anchor/priority assessment.

W4: Figure 1 evidence. A4: Need 1 (daughter's independence) follows “I'm worried she won't manage on her own” and eases with “I feel reassured she can manage on her own.” Later, “But imagining dinner without her still leaves me feeling empty” signals Need 2 (separation/loneliness) and a new stage. We will separate the evidence chains.

W5 & Q1: ETV consistency. A5: Tables 4–5 use Eq. (13), not Eq. (12). They differ: for 0→1→0, Eq. (13) gives ETV=0.5, but Eq. (12) gives 0 (opposing changes cancel). Eq. (12) captures net change; Eq. (13) weights steps by initial emotion and does not establish net recovery. We will correct the claim and recompute Eq. (12) results/rankings.

W6 & Q2–Q3: Independent validation, candidate rejection, and priority shifts. A6: We freeze K_N=3, K_A=3, W=2 selected on the original 150 cases, then evaluate 150 independent held-out cases.

Table R3. Parameter-selection vs. held-out agreement (%) and Fleiss' κ.

DimensionSelectionHeld-outκ
Need Candidate81.680.80.76
Focus Anchor83.482.70.74
Transition Decision85.984.10.80

None of the Above permits rejection of all candidates.

Table R4. Candidate coverage/rejection (%) and Fleiss' κ.

DimensionCoverageRejectionκ
Current User Need92.47.60.79
Focus Anchor89.710.30.77

(1) Satisfaction-driven Transition: update anchor after sufficient support and a shift to another unresolved concern. (2) Priority-driven Override: update it when evidence and feedback confirm a new top-priority concern despite unmet prior needs; retain the prior need/anchor in long-term memory for later confirmation if reprioritized. We annotate 200 independent dialogues (8–20 turns; 100 S/100 O) with three annotators and expert adjudication.

Table R5. Transition precision/recall/F1 (%) and response RT (s).

MethodS: P/R/F1O: P/R/F1RT S/O
Baseline45.1/46.8/45.951.5/53.4/52.40.9/1.2
UKA68.9/75.7/72.173.6/72.1/72.82.2/1.8
w/o focus anchor72.5/73.2/72.869.1/73.5/71.21.0/1.5
w/o Override76.0/79.5/77.772.4/75.8/74.11.4/1.3
NeedTracer (Full)81.7/83.4/82.580.3/82.9/81.61.2/1.2

Full achieves 82.5%/81.6% S/O-F1; O-F1 exceeds w/o Override by 7.5 pp. RT is response latency, not detection delay; memory-based resumption remains untested. We will clarify Override in Algorithm 1.

W7: Symbol errors. A7: We will fix escaped placeholders <candidate 1>, <candidate 2>, <candidate 3> in Table 7 and corresponding entries in Tables 16, 17, 19.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.