acceptodds
Under review as a conference paper at ICLR 2027

Retrospective Affective Belief Revision: When New Evidence Changes How LLM Agents Interpret Past Emotions

Abstract

An agent can recall an interaction accurately and still misunderstand how a person felt. Later disclosures can change the meaning of an earlier silence, hesitation, or emotional response, creating a need to revisit past judgments. To address this issue, we introduce Retrospective Affective Belief Revision (RABR), a framework that connects newly available evidence to historical affective interpretations while preserving the original records and the history of belief changes. A compact Trigger determines when memory retrieval is needed, a reasoning agent grounds the relevant historical target, and an evidence-based updater distinguishes between revising a judgment and suspending its application. To study when such retrospection is warranted, we construct RABRBench, spanning compact controlled situations, extended histories, and source-preserving longitudinal conversations. Those tracks pair retrieval decisions with evidence-turn annotations and dependency-aware evaluation units, making selective access to the past an explicit learning and evaluation task. Integrated experiments show that selective retrieval reduces downstream model calls, while controlled correction studies demonstrate persistent evidence support and targeted evidence retraction. Together, RABR and RABRBench establish a framework and benchmark for studying when to revisit past emotional judgments and how to revise them through traceable evidence updates.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Long-Term Memory, Affective Computing, Belief Revision, Emotion Understanding, Selective Memory Retrieval

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 2 reviews, above 27% of the others.
Reviewer_7yeDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces RABR, a framework that enables language model agents to reinterpret users’ past affective states when new evidence becomes available, while preserving original records, prior judgments, and their supporting evidence. It combines selective memory retrieval, evidence organization, and versioned belief updates, and introduces RABRBench to evaluate when to revisit history and how to revise past affective judgments.

Critical strengths
  1. The research addresses a problem of practical significance, with a framework designed around clear, focused objectives. By separating raw records from revisable sentiment interpretations—and distinguishing between "changing a judgment" and "suspending the use of a judgment"—the framework facilitates traceable, long-term user modeling.

  2. The paper introduces RABRBench, which encompasses three complementary scenarios: controlled short-context interactions, extended historical sequences, and long-context settings that preserve original dialogues. It provides turn-level annotations for the evidence used in retrieval, while accounting for dependencies between related samples during data partitioning. Furthermore, the experiments analyze the impact of maintaining versus revising judgments, as well as the effects of different retrieval methods and retrieval depths.

Critical weaknesses
  1. Much of the main text describes specific procedures and update rules, which seem closer to an operational task definition or a reference baseline; the methodological contribution remains unclear. More space should be devoted to benchmark construction, representative examples, task dimensions, and the practical importance of retrospective affective reinterpretation. A particularly important issue is whether revising past affective judgments affects subsequent answer quality or interaction behavior. The current evaluation mainly measures affect labels and revision decisions, without directly testing these downstream effects, making the practical value of retrospective revision difficult to assess.

  2. The equations and figures could be clearer. The method introduces many symbols, states, and thresholds, but their connections to task inputs, outputs, and concrete examples are not sufficiently intuitive. The figures could better illustrate how new evidence changes a historical judgment and how that change affects subsequent reasoning, helping readers understand the role of each design choice.

Questions for authors
  1. Could the authors provide representative benchmark examples and explain the main task dimensions and their practical relevance?
  2. Does revising past affective judgments affect subsequent answers or interaction behavior compared with retaining the original judgments? Are there evaluation tasks that directly measure this effect? Does the paper include tasks that require using these revised judgments to be completed correctly?
  3. Could the authors walk through a concrete example showing how the inputs, evidence scores, update rules, and resulting belief states correspond to the equations and figures?
3 comments

We thank the reviewer for the positive assessment of the problem and of RABRBench, and for the constructive comments. We address each of your concerns below.

W1a. The contribution.

RABR makes an agent's judgment of a past emotional moment a versioned, evidence-attributed object. New evidence is scored per target and per label. A fixed rule decides keep or revise, and active or suspended. Every version keeps the evidence that produced it.

The payoff is measurable. On identical history, this reaches 0.996 final-affect accuracy on LC (uniform root; 0.872 in the submitted configuration). Direct reclassification reaches 0.743, and prompting the model to reconsider its previous judgment 0.782. Because the trail is stored, every REVISE miss in Table 3 could be traced to one cause and corrected without re-running a model. In the revision, Section 3 will be shortened and the space given to the benchmark.

Q1. Representative cases and task dimensions.

Each LC case gives the agent a current conversation, 12-49 memory items from earlier sessions, and a standing judgment about one earlier moment.

  1. LoCoMo conv-26 (source-preserved history, native session boundary). Target, Caroline: "...the support was the best! Such a special moment." Judgment: JOY. Later: "...at that exact time I actually felt a heavy sadness." Gold: REVISE to SADNESS. Walked through in full under Q3.
  2. AMI IS1004 (meeting, controlled insertion). Target, Project Manager: "Come on, guys, we've only got three minutes, I need your inputs now." Judgment: ANGER, from an indirect cue. Later: "I wasn't annoyed at you. I was scared we'd run out of time." Gold: REVISE to FEAR. 49 items, target at position 35.
  3. AMI ES2005 (KEEP control). Target: "Oh no, this is so embarrassing... I've completely blanked on it." Judgment: OTHER. Later: "I was genuinely embarrassed, not just irritated." Gold: KEEP.
  4. MSC test_44. Target: "I'm scared of horses. Too big for me." Judgment: FEAR. Later: "honestly I was mostly embarrassed... so scared was easier to say." Gold: REVISE to OTHER, the hardest category for every method.

The task dimensions are:

  • (a) Retrieval demand: request vs stop, with annotated cue turns (SC, EC).
  • (b) Target localization among many candidates (LC): a Jaccard retriever ranks the target first in only 0.58 of cases.
  • (c) Keep vs revise: 246/221 in the LC test split.
  • (d) The replacement label: 8 categories.
  • (e) Evidence type: confirming vs overturning, direct vs indirect cue, natural (Route A, 27 cases) vs controlled (Route B, 440).

(To be continued)

Q3 / W2. One example through Figure 2 and every equation.

We follow case 1 through Figure 2, panel by panel. The data is the case as stored in RABRBench, and all numbers come from the stored trace of the submitted run.

Input. The agent holds 17 memory items from an earlier session and a judgment about one of them. It then reads the 25 turns of a new session, one at a time.

# Persistent memory: earlier session (17 items; items 9-11 shown)
{memory_item:  9, speaker: "Melanie",  text: "Wow! What's the best part you remember from it?"}
{memory_item: 10, speaker: "Caroline", text: "Seeing my mentee's face light up when they saw the support was the best! Such a special moment.",
                  historical_target: true, initial_affect: "JOY", cue_channel: "INDIRECT_CONTEXTUAL_CUE"}
{memory_item: 11, speaker: "Melanie",  text: "Wow, Caroline! They must have felt so appreciated. It's awesome to see the difference we can make in each other's lives. ..."}

# Current session (25 turns; turns 1-2 and 23-25 shown)
{turn:  1, speaker: "Caroline", text: "Hey Melanie! Just wanted to say hi!"}
{turn:  2, speaker: "Melanie",  text: "Hey Caroline! Good to talk to you again. What's up? Anything new since last time?"}
...
{turn: 23, speaker: "Caroline", text: "Wow, Melanie, what a beautiful moment! Lucky you to have such an awesome family!"}
{turn: 24, speaker: "Melanie",  text: "Thanks, Caroline! I'm really lucky to have my family; they bring so much joy and love."}
{turn: 25, speaker: "Caroline", text: "Mel, when I told you about seeing my mentee's face light up at the pride event, I called it the best and a special moment—but at that exact time I actually felt a heavy sadness. It brought back how alone I felt as a teen. I just didn't say it then.",
           delayed_evidence: true, referenced_target: memory_item 10}

# Reference
{action: "REVISE", initial_affect: "JOY", final_affect: "SADNESS", turns_from_target_to_history_end: 7}

The fields historical_target, initial_affect, delayed_evidence, referenced_target and the reference are annotations. The agent sees only speakers and text, plus its own judgment (JOY for item 10).

Panel by panel.

Figure 2 panelIn plain wordsSymbol / equationValues in this case
Persistent memory B (before)The agent already holds a judgment about the earlier moment.Eq. 1: B_h = (T, p, c, r)c = JOY. p = 0.514 for JOY, 0.069 for each other label. r = ACTIVE.
01 Current observationThe agent sees only the current turns.I_tTurns 1-25 of the current session, one at a time.
02 TriggerA small model reads the current turns and decides whether to look back.Eq. 2: d_t = pi(I_t), q_tTurns 1-24: STOP. Turn 25: REQUEST MEMORY, with selected turn q_t = {turn 25}.
03 Query & retrieveThe selected turn becomes a query. Candidate records are read from memory.Eq. 2: Compile(q_t), M_tFull history: all 17 earlier items. BM25 variant: items 10, 6, 17. The target (item 10) is ranked first.
04 Reasoning agentThe agent finds which earlier moment the turn is about, and scores every label for it.Ground target h ("Existing target")h = Caroline's "special moment". SADNESS: STRONG_FOR (+2). JOY: WEAK_AGAINST (-1). Other six labels: 0.
05 Evidence admissionThe four checks decide whether this evidence may be used.E_gTurn 25 passes all four checks and forms one dependency group. Item 10 is the target's own record, not new evidence, so it is not admitted.
06 Reconstruct affective supportStarting score plus the evidence, per label.Eq. 3: z = b + s, p = softmax(z)See the small table below.
07 RevisionIs there a new leader, and is it clearly ahead?Eq. 4, commitment cLeader SADNESS differs from JOY. Shift from anchor 0.389 >= delta (0). Lead over JOY 0.290 >= gamma (0). Result: REVISE.
07 ResolutionIs the leader unique and strong enough to act on?Eq. 5, resolution rPeak 0.459 >= 0.125. Margin 0.290 >= 0. Result: ACTIVE. The agent now shows SADNESS.
"Append version", Persistent memory B (after)A new version is saved with a parent link. The old one is kept.Eq. 6: B_h(tau; T)The new version lists turn 25 in its evidence manifest. Asked "what did you believe before turn 25?", the agent answers JOY. Asked now, it answers SADNESS.

(To be continued)

Panel 06 in detail. Evidence scores are centered (their mean, 0.125, is subtracted), so evidence only moves labels relative to each other. This is the bar chart in panel 06.

LabelStarting score bEvidence score (raw)Evidence score (centered) sTotal z = b + sShare p = softmax(z)
JOY (old judgment)2-1-1.1250.8750.169
SADNESS0+2+1.8751.8750.459
Each of the other six00-0.125-0.1250.062

The same case ending in SUSPENDED. Had the agent scored JOY as NEUTRAL (0) instead of WEAK_AGAINST, the mean would be 0.25. JOY would total 2 - 0.25 = 1.75, and SADNESS 0 + 1.75 = 1.75: an exact tie. Panel 07 then finds no unique leader. The commitment stays JOY, the resolution is SUSPENDED, and the agent shows UNKNOWN, as at the bottom of panel 07. This pattern accounts for 52 of the 59 REVISE misses of RABR (Full history) in Table 3. Without the +2 on the old label (uniform root), the same stored scores give 220/221.

We will add this walk-through and its two tables in the revised version.

W1b / Q2. Revision change later behaviour.

To our best knowledge, no existing benchmark has a task that can be completed only with the revised judgment. We checked 32 multi-session memory and emotional-support datasets: their updates are fact overwrites or preference drift, not reinterpretation of an earlier moment. The paper's case for revision rests on the framing in Sections 1-2 (belief revision, truth maintenance, appraisal-based reinterpretation) and on the applications it names. We have now measured the effect directly, in two ways.

Both use the 467 strict LC cases and the same generator (Qwen3.8-27B-FP8, greedy), conditioned on one of four stored beliefs about the earlier moment: the fixed initial judgment, Full-history Reclassification's output, RABR (Full history)'s output, and gold.

Free reply. (A) The agent replies right after the disclosure. (B) In a later session the person says "I was thinking again about when I told you: ... How do you remember that moment for me?", with the disclosure out of context, so only the stored belief carries the reinterpretation. A blind DeepSeek judge (temperature 0, conditions hidden, order randomized) scores 1-5: consistency with the person's actual feeling (R1), no contradiction with the disclosure (R2), appropriateness (R3). 3,736 absolute and 863 pairwise judgements, no parse failures.

Situation B, gold REVISE (n=221)R1R2R3
Fixed initial judgment1.041.061.66
Full-history Reclassification output3.864.063.80
RABR (Full history) output3.853.923.70
Gold label (ceiling)4.604.594.15
  • RABR minus fixed initial on R1: +2.81 (95% family-cluster CI [+2.52, +2.99]). Pairwise preference for the RABR reply: 0.897 (192/214).
  • For case 1, the fixed-initial agent says: "I remember that moment as a burst of pure joy for you." The RABR agent says: "a time of deep sadness for you, even though your words described such joy."
  • RABR's 169 committed revised beliefs score 4.65, level with gold. Its 52 suspended beliefs score 1.25: the tie case above, seen downstream.
  • KEEP: the two agents' replies are identical on 245/246, so revision does not harm retained judgments.
  • Situation A, disclosure in view: all conditions 4.2-4.3. The stored belief matters once the disclosure has left the context.

Exact-match task. Same cases and beliefs, later session, disclosure out of context, fixed answer options. Q: how did the person feel then (8 labels + unsure). S: which of four replies to send (gold-consistent, initial-consistent, other label, hedge). R: "lately I have been feeling X; does that earlier moment count as a time I felt the same?" (yes / no / unsure).

REVISE accuracy (n=221)QSR
Fixed initial judgment0.0050.0000.000
RABR (Full history) output0.7560.7380.744
RABR, committed stratum (169)0.9590.9590.947
Gold label0.9951.0000.989

RABR minus fixed initial on Q: +0.751 (CI [+0.673, +0.799]). KEEP: 0.996-0.998 for both agents. The generator adopts a committed stored label in 169/169 cases, so downstream accuracy equals belief accuracy. Without the revised judgment the task cannot be completed.

Both evaluations go into Section 5. Their scope (single generator, templated items, construction-time gold in 440/467 cases) will be stated in Limitations.

W2. Presentation.

The walk-through under Q3 ties each symbol, threshold and panel of Figure 2 to a concrete input and output, and will be added to the paper. Table 1's columns will name the task dimensions directly.

We thank the reviewer again for the careful review. We look forward to your response and are happy to clarify any remaining questions.

Reviewer_W3yeOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RABR, a framework that enables LLM agents to revisit and revise their interpretations of past affective states when new evidence becomes available. RABR separates historical event records from revisable affective beliefs and incorporates selective memory retrieval, evidence-based belief updates, and versioned memory. The authors also introduce RABRBench, a benchmark spanning controlled scenarios and long conversational histories to evaluate retrieval decisions and retrospective affective revision. results show improvements in affective revision performance on long-context tasks and reduced computational costs through selective memory retrieval.

Critical strengths

The paper distinguishes historical event records from revisable affective interpretations, allowing agents to update past judgments while preserving their original beliefs and supporting evidence. This provides a potentially useful approach to maintaining interpretable long-term memory.

RABRBench evaluates both when an agent should retrieve historical information and how it should revise previous affective judgments. Its combination of controlled scenarios and longer conversational histories provides a structured setting for studying retrospective reasoning.

Critical weaknesses

Insufficient motivation and limited novelty. While the idea of retrospectively revising affective beliefs is interesting, the paper does not sufficiently establish why this capability is important for long-term conversational agents. In particular, it remains unclear what practical problems arise when agents fail to revise past affective interpretations, or how such revisions improve subsequent interactions. Furthermore, separating factual records from revisable beliefs is not an entirely new idea, making the novelty of the proposed framework less convincing.

Limited insights from experimental results. The main findings appear somewhat expected. For example, the conclusion that agents need to identify when previous judgments should be revised and correctly determine the updated affective state largely follows from the task formulation itself. The paper would benefit from more fine-grained analyses that reveal less obvious challenges in retrospective affective reasoning, such as which types of evidence make revision particularly difficult and why models fail to identify the correct affective state even when relevant information is available.

Insufficient ablation of the proposed framework. RABR introduces several components, including evidence grouping, anchor-based revision, suspension, and versioned memory. However, the experiments do not sufficiently isolate the contributions of these components. It remains unclear whether such a complex framework is necessary or whether simpler approaches, such as directly prompting an LLM to reconsider its previous judgment using new evidence, could achieve comparable performance.

Questions for authors
  1. can the authors please provide additional evidence that retrospective affective belief revision improves downstream conversational behavior or long-term user modeling, beyond improving the accuracy of historical affective labels?
  2. Which components of RABR contribute most to its performance, and how does the proposed framework compare with a simpler prompting-based belief revision approach?
8 comments

We thank the reviewer for the positive assessment of RABR and RABRBench, and for the constructive comments. We address each of your concerns below.

W1a. Motivation: what breaks when an agent does not revise.

The motivation rests on established theory. Sections 1-2 introduce belief revision (AGM) and truth maintenance (Doyle, 1979, cited in the paper): a belief must be revisited when its supporting evidence changes. Work in psychology and cognitive science shows why this applies to judgments about emotion:

  • An inferred emotion is not the felt emotion. Surface signals do not reliably reveal what a person feels [1]. For text, human and LLM annotations diverge from authors' own reports (Li et al., 2025, cited in the paper).
  • Observers infer emotion from a person's beliefs, goals and appraisals. Formal models describe this as Bayesian inference over the person's mental states [2]. New information about those states, such as a later disclosure, should therefore change the inferred emotion.
  • The emotional meaning of a past moment changes over time. Remembering an emotional event is a reconstruction that can change the memory itself [3], and reinterpreting a situation changes its emotional meaning [4].

An agent that keeps its first reading of a past moment ignores all three. Player modelling, affect-sensitive tutoring and emotional support all depend on getting this right.

To our best knowledge, this is not yet a complete task in NLP. No existing benchmark tests whether an agent reinterprets a past emotional moment and then acts on the new reading. We checked 32 multi-session memory and emotional-support datasets: their updates are fact overwrites or preference drift. This is why the submission could not show the practical effect on an existing benchmark, and why RABRBench is needed.

We therefore measured the effect directly (details under Q1). In both measurements the agent is in a later session, the person mentions the earlier moment again, and the disclosure is no longer in context. Only the stored judgment differs. On the 221 cases that require revision:

  • Free reply, blind judge (1-5). Fixed initial judgment 1.04 (213/221 cases scored 1). RABR's revised judgment 3.85. Gold 4.60. Pairwise preference for the RABR reply 0.897.
  • Exact-match task. Fixed initial judgment answers 0.000-0.005 of items correctly. RABR answers 0.738-0.756. Gold answers 0.989-1.000.

A concrete case (LoCoMo conv-26). Caroline once called watching her mentee at a pride event "the best, such a special moment". Sessions later she says she actually felt heavy sadness, because it recalled how alone she was as a teen. Later, the non-revising agent says: "I remember that moment as a burst of pure joy for you." The revising agent says: "a time of deep sadness for you, even though your words described such joy."

W1b. Novelty.

We agree, and the paper says, that separating facts from beliefs goes back to truth maintenance (Doyle, 1979) and temporal databases. That is not our claim. Our claim is the mechanism that makes affective judgments revisable in a long-term agent. None of the cited systems has it:

  • A-MEM updates linked notes, but does not weigh evidence against a specific person-episode judgment.
  • Memory-R1 learns add, update and delete over memories, not whether the reading of a past moment should change.
  • Zep keeps a bitemporal graph of facts, not of judgments that can be suspended.
  • LongMemEval tests knowledge updates, not reinterpretation.

RABR contributes (i) target-relative evidence applications, (ii) revise and suspend as distinct outcomes, (iii) exact retraction from the evidence trail, (iv) as-of queries over affective beliefs, and (v) the first benchmark pairing when-to-retrieve with keep/revise supervision for affect. On identical history this gives 0.996 against 0.743 for direct reclassification (W3 / Q2). Every analysis in this rebuttal, from the tie diagnosis to the downstream study, needed no new model run, because the evidence trail is stored. That is the point of the design.

Refs

[1] Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements. Psychological Science in the Public Interest, 20(1), 1-68. https://doi.org/10.1177/1529100619832930

[2] Ong, D. C., Zaki, J., & Goodman, N. D. (2019). Computational models of emotion inference in theory of mind: A review and roadmap. Topics in Cognitive Science, 11(2), 338-357. https://doi.org/10.1111/tops.12371

[3] Kensinger, E. A., & Ford, J. H. (2020). Retrieval of emotional events from memory. Annual Review of Psychology, 71, 251-272. https://doi.org/10.1146/annurev-psych-010419-051123

[4] Uusberg, A., Taxer, J. L., Yih, J., Uusberg, H., & Gross, J. J. (2019). Reappraising reappraisal. Emotion Review, 11(4), 267-282. https://doi.org/10.1177/1754073919862617

(To be continued)

W2. Which cases are hard, and why models fail with the evidence in hand.

Errors on the 467 strict cases:

MethodKEEP errors (246)REVISE errors (221)
Full-history Reclassification52 abstain, 6 wrong label57 abstain, 5 wrong label
Mem067 abstain, 9 wrong label49 abstain, 1 keep initial, 3 wrong label
RABR (Full history)1 suspended52 suspended (exact ties), 7 keep initial
RABR (Retrieved history)055 suspended, 32 keep initial (27 with no application returned), 1 failure
RABR (tuned)7 wrong label57 keep initial, 23 wrong label
  1. Prompted readouts abstain rather than mislabel (109/120 and 116/129 errors). They cannot bind the evidence to the right historical target.
  2. Deterministic RABR fails by suspension. In all 52 and 55 cases the agent cited the delayed turn, scored the gold label STRONG_FOR and the held label NEUTRAL. After centering, this exactly cancels the +2 that LC's root profile puts on the held label, giving a tie. With a uniform root, the same stored scores resolve 220/221 (Full) and 192/221 (Retrieved) REVISE cases. The failure lies in how much one disclosure must outweigh the held judgment, not in reading the evidence.
  3. The learned updater under-revises. 57 of its 80 misses keep the initial label.

Hard strata. Gold OTHER is hardest: RABR 10/26, Reclassification 5/26, tuned 0/26; NEUTRAL to OTHER is 0/6 for every method. Natural source-grounded targets (Route A, 15 REVISE cases) score 0.53-0.60 against 0.73-0.77 on controlled cases, for the three strongest methods. Initial judgments resting on a direct lexical cue are harder for RABR to revise (0.59 vs 0.79), consistent with the prior lean. Distance from target to end of history has no monotone effect.

(To be continued)

W3 / Q2. Component ablation and a simpler prompting baseline.

Ablation. The agent's scores are stored, so we replay only the updater with one part changed at a time, with no new model calls. The unchanged replay reproduces Table 2 exactly. RABR (Full history), correct cases out of 467:

VariantAll
Uniform root (no +2 on the held label)465 (REVISE 220/221)
No dependency grouping407
Admit history-only citations377
No suspension408
Agent scores only, no updater465

Accuracy comes from the agent's target-specific scoring of evidence. The updater supplies traceability, retraction and the revise/suspend split. Its prior lean is the one setting that trades REVISE recall for KEEP stability.

Prompting baseline. This exists as a retained run on the same 483 LC cases, with the same Qwen3.8-27B-FP8 at temperature 0. The prompt shows the previous judgment and asks the model to keep or revise it given the evidence. Rescored on the 467 strict references:

Prompted reconsideration with ...KEEPREVISEAll
current observation only0.5930.4070.505
BM25 top-3 items0.9470.7010.831
the oracle target item0.9800.6520.824
the full memory pool0.9020.6470.782
RABR (Full history): Table 20.9960.7330.872

With the same full history, prompting gives 0.782 and RABR 0.872. The prompted model must locate the target, weigh the evidence and decide in one generation. RABR asks only for per-target, per-label evidence scores and leaves the decision to a rule whose inputs are all stored. The prompt and harness differ from Table 2; model, decoding, cases and references are the same.

Q1. Downstream behaviour and user modelling.

Both measurements from W1a use the 467 strict LC cases and four stored beliefs about the earlier moment: fixed initial; Full-history Reclassification output (109 abstentions shown as "uncertain"); RABR (Full history) output (52 suspensions shown as "under review"); gold. The generator is Qwen3.8-27B-FP8, greedy.

Free reply. Two situations: (A) reply right after the disclosure; (B) reply in a later session when the person re-mentions the moment, with the disclosure out of context. A blind DeepSeek judge (temperature 0, conditions hidden, order randomized) scores 1-5: consistency with the person's actual feeling (R1), no contradiction with the disclosure (R2), appropriateness (R3). 3,736 absolute and 863 pairwise judgements, no parse failures.

Situation B, gold REVISE (n=221)R1R2R3
Fixed initial judgment1.041.061.66
Full-history Reclassification output3.864.063.80
RABR (Full history) output3.853.923.70
Gold label (ceiling)4.604.594.15
  • RABR minus fixed initial: R1 +2.81 (95% family-cluster CI [+2.52, +2.99]), R2 +2.86, R3 +2.04. Pairwise preference for RABR 0.897 (192/214; 2 the other way).
  • RABR's 169 committed revised beliefs score 4.65 on R1, level with gold. Its 52 suspended beliefs score 1.25: the tie from W2, seen downstream.
  • KEEP: RABR and fixed-initial replies are identical on 245/246. Reclassification's abstentions lower KEEP R1 by 0.20.
  • Situation A, disclosure in view: all conditions 4.24-4.33, pairwise 0.51. The stored belief matters once the disclosure has left the context, which is the long-term setting the paper targets.

Exact-match task. Same cases and beliefs, later session, disclosure out of context. Q: how did the person feel then. S: which of four replies to send. R: does that moment count as a time the person felt X.

REVISE accuracy (n=221)QSR
Fixed initial judgment0.0050.0000.000
RABR (Full history) output0.7560.7380.744
RABR, committed stratum (169)0.9590.9590.947
Gold label0.9951.0000.989

KEEP is unaffected (0.996-0.998 for both agents). The generator adopts a committed stored label in 169/169 cases, so downstream accuracy equals belief accuracy. Without the revised judgment, the task cannot be completed. Both evaluations go into Section 5, with prompts, rubric and tables in the appendix.

On "the findings follow from the task formulation". The content is where deciding to revise and choosing the label come apart. The two deterministic RABR variants reach 0.93-0.98 revision agreement, but 0.60-0.73 REVISE final-label accuracy. The ablation shows the gap is one prior setting. The paper's claims will be stated at exactly this level.

We thank the reviewer again for the careful review. We look forward to your response and are happy to clarify any remaining questions.

Comments on the paper

3 comments

We thank the reviewer for the positive assessment of our work and for the constructive comments. We address each of your concerns below.

All numbers use the Table 2 population: 467 strict LC test cases (246 KEEP, 221 REVISE). Failures and unresolved outputs count as wrong.

W1 / Q2. The REVISE gain is not significant.

Family-cluster bootstrap over the 13 LC source families (10,000 resamples) and exact sign test, RABR (Full history) minus Full-history Reclassification:

Subsetngains/lossesdiff95% cluster CIsign p
All46790/30+0.128[+0.090, +0.163]3.8e-8
KEEP24657/0+0.232[+0.171, +0.288]1.4e-17
REVISE22133/30+0.014[-0.034, +0.054]0.80

The cause is one prior setting. All 52 REVISE misses of RABR (Full history), and 55 of RABR (Retrieved history), are exact ties. In each case the agent found the disclosure: it cites the delayed-evidence turn in 221/221 REVISE cases. It scored the gold label STRONG_FOR (+2) and the held label NEUTRAL. After centering, the gold label gets +1.75 and the held label -0.25. LC's root profile adds +2 to the held label, so both end at 1.75. With no unique leader, Eq. 5 suspends the judgment and it scores zero.

Appendix A will state the root profiles exactly: +2 on the held label for LC, uniform for SC and EC, which have no initial label.

Q3. Component ablation.

Setup. The reasoning agent scores every label for the target (Figure 2, panel 04). The updater is a fixed rule that turns these scores into a decision through four parts:

  • Admission (panel 05): only evidence new at this step is counted.
  • Dependency grouping (Eq. 3, the 1/n_g term): applications from one source are averaged.
  • Root profile b_h (Eq. 3): the starting score of each label. On LC, +2 on the held label.
  • Suspension (Eq. 5): with no clear leader, the judgment is SUSPENDED and shown as "unknown".

The agent's scores are stored. We keep them fixed and replay only the updater, changing one part at a time. No model calls are needed, and any change comes from that part alone. The unchanged replay reproduces Table 2 exactly: 407/467 for Full history (KEEP 245, REVISE 162) and 379/467 for Retrieved history.

Results. Correct cases for RABR (Full history), out of 246 KEEP, 221 REVISE and 467 All.

VariantPart changedKEEPREVISEAll
Uniform rootroot profile: 0 for every label245220465 (0.996)
No dependency groupinggrouping: every application counted on its own245162407 (0.872)
Admit history-only citationsadmission: the target's own earlier record also counted245132377 (0.807)
No suspensionsuspension: never SUSPENDED246162408 (0.874)
Agent scores only, no updaterupdater removed: top label of the summed scores if positive, else the held label245220465 (0.996)
  1. Uniform root: +58 REVISE. This removes the tie described under W1 / Q2.
  2. No grouping: no change. Every LC source gives exactly one application, so nothing is averaged.
  3. History-only citations: -30 REVISE. The target's own record produced the old judgment. Counting it again pulls the decision back, so the admission rule is needed.
  4. No suspension: +1 KEEP only. A tie still has no new leader, so the old label is kept instead of "unknown". The errors come from the tie, not from suspension.
  5. Agent scores only: same as the uniform root. Accuracy comes from the agent's target-specific scoring. The updater adds a stored, traceable and retractable decision, and the suspend state.

Size of the lean. Below 2: 459-465. Exactly 2: 407 (tie). 2.1 to 2.5: 408 (the old label wins). 3.0: 292. Retrieved history with the uniform root: 438/467 (REVISE 192/221).

This replay was computed on the test split. We report it as a sensitivity analysis next to Table 2, not as a replacement.

(To be continued)

W2 / Q1. Current observation only, with matched model and decoding.

Setup. A retained run on the same 483 LC test cases, with the same Qwen3.8-27B-FP8 at temperature 0. The prompt shows the earlier judgment and asks for KEEP or REVISE plus a final label. Only the history given to the model varies. The run is therefore also the "prompt the model to reconsider" baseline. We rescored it on the 467 strict references.

Rows 1-4 are this run; rows 5-6 are from Table 2. Cells are final-label accuracy. "Unresolved = wrong" is the Table 2 rule. The last column instead counts an abstention or suspension as keeping the earlier judgment.

History given to the modelKEEPREVISEAll (unresolved = wrong)All (unresolved = keep earlier judgment)
None: current observation only0.5930.4070.5050.719
BM25 top-3 memory items0.9470.7010.8310.839
Oracle: only the target memory item0.9800.6520.8240.824
Full memory pool0.9020.6470.7820.797
Table 2: Full-history Reclassification0.7640.7190.7430.854
Table 2: RABR (Full history)0.9960.7330.8720.874
  1. No history. The model abstains on 219/467 cases, saying it cannot tell which earlier message the judgment refers to. It gets 0.407 of REVISE cases from explicit cues in the current turn.
  2. Any history. All accuracy rises from 0.505 to 0.78-0.83, and REVISE from 0.407 to 0.65-0.70.
  3. Same full history. Asking the model to reconsider gives 0.782. RABR gives 0.872, and 0.996 with the uniform root.

Rows 1-4 use a prompt that shows the earlier judgment, unlike Table 2. Model, decoding, cases and references are the same. The table will go into the appendix with this note.

Lexical cues. Appendix C reports a word naming the final label in 86-88% of delayed-evidence turns. Row 1 shows what these words give without history: 0.407 on REVISE. On the 26 REVISE turns with no such word anywhere in the scene (Appendix J.2), RABR gains five exact final labels and loses one relative to Reclassification.

Summary. The REVISE gain of the submitted configuration is not significant. Its cause is one prior setting, found from the stored evidence trail and reversible without re-running the model. A matched current-only baseline isolates what history adds: 0.505 without it, 0.872 for RABR with it. We will add all three analyses to the paper.

We thank the reviewer again for the careful review. We look forward to your response and are happy to clarify any remaining questions.

Thank you for the responses and clarifications. The rebuttal has largely addressed my concerns. I also consider the concerns raised by the other reviewers important, so I will maintain my current score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.