acceptodds
Under review as a conference paper at ICLR 2027

Mind the Supersession Gap: Diagnosing Event-to-Preference Inference Failures in Language Models

Abstract

Alignment between a user and a personalized assistant is not a static achievement. Circumstances evolve over time, and users typically communicate these shifts through life events rather than explicit declarations. Whether a preference revision successfully registers with the assistant is therefore a property of the coupled interaction system rather than a capability of the model in isolation. We introduce a novel diagnostic framework that rigorously isolates implicit event-to-preference inference from standard memory retrieval. Across an evaluation of nine open-weights models and four API models, we find that assistants honor explicitly declared revisions with near-perfect accuracy but rarely recognize revisions implied by events. Performance falls significantly below chance at 2.2 to 22.1 percent. These implicit signals are nevertheless highly legible to humans. Annotators resolve the same updates at 100 percent accuracy on items where they endorse the premise. We demonstrate that this deficit is neither a long-context retrieval failure nor a bottleneck solvable by increased test-time computation. Blind-scored reasoning traces reveal a dissociation between stated reasoning and final choice. Models that derive an event's consequence are no more likely to act on it than those that do not. Mechanistic analysis using the Jacobian lens shows that the event-implied value emerges early but loses influence by the answer. Contrastive steering and activation patching can recover implicit updates in some settings, but no intervention tested provides a general repair because increasing uptake of the inferred preference can also weaken adherence to explicitly stated preferences. The strongest predictor we identify is the weight the interaction places on the initial statement, which is determined by the framing. Recasting the exchange as an ordinary conversation improves accuracy by 52 points, while inducing a mirror vulnerability where assistants improperly override explicit preferences. A deployed system can thus be overly deferential to initial statements and insufficiently deferential to recent ones. We have provided the datasets, model output, and code.

Primary area:
interpretability and explainable AI
Keywords:
human-AI alignment, dynamic alignment, personalization, preference updating, LLM agents, instruction following, behavioral evaluation, chain-of-thought faithfulness

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_a8FfOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines whether assistants revise their recommendations when a subsequent life event makes an earlier preference inappropriate. It introduces synthetic conversations that distinguish unchanged preferences, explicit revisions, and revisions implied by events. Across nine open-weight models and four API models, the authors find a substantial gap between following explicit updates and inferring implicit ones under the default examination framing. They further examine how this gap varies with context length, deliberation, and conversational framing. Reasoning-trace analysis and Jacobian/R-lens readouts suggest that information supporting an alternative option may be present without determining the final answer. Steering improves implicit updating but can severely impair retention of unchanged preferences.

Critical strengths
  • The explicit-versus-implicit comparison provides a useful diagnostic. T0 and T1 test whether models can retain an existing preference and follow an explicitly stated revision, respectively. T2 then tests whether they can infer a revision from an event. The T5 minimal pair helps isolate the effect of including the original preference among the answer options. The gap appears across multiple model families and in both probing and generation settings, making the finding more convincing than a result limited to a single model or response format.

  • The framing and steering experiments reveal tradeoffs that would be missed by evaluating implicit-update accuracy alone. Conversational framing improves T2 performance but introduces errors in following explicitly stated preferences. Steering likewise improves T2 while severely degrading T0. These results show why improving performance in one condition does not necessarily yield a better preference-following system overall.

Critical weaknesses
  • The human validation does not fully resolve this concern. Table 17 reports 75% agreement with the answer key on 20 T2 items evaluated by one annotator. The reported 100% ceiling is based on the 15-item subset for which that annotator endorsed the premise; the stricter subset contains only 10 items. These results support the existence of clear model failures on some examples, but leave the validity of the benchmark’s labels more broadly uncertain. Independent validation, with results reported separately for push and pull events, would strengthen the central claim.

  • The steering results also appear consistent with suppressing the original preference, given the collapse in T0 accuracy. They therefore provide limited evidence that the intervention selectively restores a correctly computed update. Finally, the abstract and Section 7 discuss activation patching, but I could not locate its experimental setup or quantitative results in the manuscript. This makes that part of the causal argument difficult to assess.

Questions for authors
  • How much of the behavioral gap remains on push-only items with independently validated premises? Please report these results separately from those for pull items. Does the human–model gap remain large when evaluation is based on agreement among independent annotators, rather than a subset selected using a single annotator’s premise judgments?

  • Can the event-token readout distinguish between events within the same attribute that support different recommendations, or between events that make an option inappropriate and those that merely mention it? Appendices C/I identify this missing control. Such comparisons would help establish whether the readout captures an updated preference rather than an association with an option.

6 comments
AuthorsOpenReview ↗

(1/4) Q2. Can the event-token readout distinguish same-attribute events that support different recommendations?

We thank the reviewer. Both questions identified missing controls, and we ran both experiments. Two new independent annotators have also completed the validation set, addressing the first weakness. The remaining weakness was a production error on our side: activation-patching results were omitted from the manuscript even though the abstract and Section 7 still refer to them. We provide the missing results below.

Our submitted control paired each T2 item with a twin whose event came from a different attribute. As the reviewer notes, this could not distinguish whether the readout tracks an updated preference or simply an association with a topic.

We reran the control using donor events from the same attribute, but implying a different one of the three preference states. Both prompts therefore concern transit or exercise and differ only in which option the event supports. The conversation, turn-1 statement, option pair, and seed remain fixed.

The effect becomes stronger, not weaker.

ModelDonorPull (n=69)Push (n=51)Pull vs. push
Gemma-3-12bCross-attribute+10.15, p=4e-8−0.71, p=0.46p=6e-6
Gemma-3-12bWithin-attribute+13.87, p=3e-9+1.49, p=0.61p=2e-4
Qwen3.5-9BCross-attribute+1.44, p=1e-4−0.04, p=0.93p=0.002
Qwen3.5-9BWithin-attribute+2.05, p=4e-8+0.31, p=0.34p=6e-4

These values are paired differences in the event-token margin against each item's twin, using the same layer band and readout as the main mechanism table. The reported p-values use a two-sided Wilcoxon signed-rank test; the final column uses a Mann-Whitney test between subgroups.

Across all 120 items, the paired difference increases from +5.53 to +8.61 on Gemma-3-12b (p=7e-7, 88/120 items positive) and from +0.81 to +1.31 on Qwen3.5-9B (p=9e-7, 83/120 items positive). The ballot effect at the answer token remains unchanged at +20.09, confirming that the rest of the pipeline is unchanged.

Matching the attribute strengthens the contrast. This supports the interpretation that the event-token readout tracks which option the event points to, rather than simply the topic.

This result also requires us to revise our explanation of the push null. We had attributed it to the donor event being topically unrelated to either option. Since matching the attribute does not resolve the push null, that explanation is no longer supported.

The remaining possibility is that the readout tracks the specific option, which the donor event does not mention. A cleaner test would pair each eliminating event with a same-topic event that does not eliminate the option, such as "I finally got the bike serviced" versus "I fractured my wrist". This would isolate elimination while holding the topic fixed. We have not built these stimuli, so we cannot yet establish this explanation. We will state this limitation explicitly rather than retain the weaker explanation.

AuthorsOpenReview ↗

(2/4) Q1 and W1. The behavioural gap on push items with independently validated premises

The main result remains. Two more independent annotators have completed the same 60 items, so we can now assess the premises using agreement between three readers rather than one reader's judgment.

The annotators agree on the answer for 91.7% of all 60 items (Cohen's κ=0.91) and 85.0% of the 20 T2 items (κ=0.84). They agree on the premise check for 96.7% of items (κ=0.93). For the clearly-better check, raw agreement is 86.7%, while κ is 0.44. This lower κ reflects the high base rate of "yes" responses. All five answer disagreements occur in T2 and T4, where the premise is contestable.

The table below reports results by event type for the new annotators. Each gate requires agreement between both annotators.

GateForcenAnnotator AAnnotator BModels
Both accept premisePull7100.0%85.7%1.0% over 104
Both accept premisePush8100.0%87.5%14.3% over 98
Both gatesPull4100.0%100.0%1.5% over 66
Both gatesPush6100.0%100.0%2.9% over 70

On the subset where both annotators accept the premise and agree that the alternative is clearly better, both answer every item correctly for both event types. Models reach only 1.5% on pull and 2.9% on push. Pooled across event types, both annotators score 100% on these 10 items, compared with 2.2% for the models. Models are evaluated on exactly the same items that pass each gate.

Every human error by either annotator selects the original stated option, despite that annotator having marked the premise as disputed. These are disagreements with the answer key, not failures to infer the implied preference.

Results across the full item set

The following table reports the T2 results from our main table, split by event type using the same hand annotation, items, and parser. The pooled column reproduces the main table. Chance is 50%.

ModelReadoutPullPushPooledp
Qwen3.5-0.8BProbe2.9 (n=276)3.2 (n=186)3.01.000
Qwen3.5-2BProbe3.3 (n=276)2.7 (n=186)3.00.789
Qwen3.5-4BProbe0.7 (n=276)4.3 (n=186)2.20.017
Qwen3.5-9BProbe2.5 (n=276)9.7 (n=186)5.40.001
Qwen2.5-7BProbe10.1 (n=276)13.4 (n=186)11.50.299
Qwen3.5-2BGen.2.7 (n=74)2.4 (n=42)2.61.000
Qwen3.5-4BGen.0.0 (n=76)6.8 (n=44)2.50.047
Qwen3.5-9BGen.1.8 (n=166)8.1 (n=74)3.80.027
Qwen2.5-7BGen.12.7 (n=166)14.9 (n=74)13.30.683
Gemma-3-1bGen.14.5 (n=76)0.0 (n=44)9.20.007
Gemma-3-4bGen.2.6 (n=76)4.7 (n=43)3.40.619
Gemma-3-12bGen.5.3 (n=76)15.9 (n=44)9.20.096
Gemma-3-27bGen.2.6 (n=76)13.6 (n=44)6.70.050
gpt-5.4-nanoGen.0.0 (n=276)0.0 (n=186)0.01.000
gpt-5.4-miniGen.2.5 (n=276)5.4 (n=186)3.70.133
gpt-5.4Gen.12.3 (n=276)15.1 (n=186)13.40.407
gpt-5.6-solGen.14.5 (n=276)33.3 (n=186)22.1<1e-5

The gap is at least as large on push as on pull for every model except Gemma-3-1b, the weakest model in the study. The highest push result is 33.3% for gpt-5.6-sol, still below chance. Eleven of the seventeen rows are below 10%.

This distinction matters for interpreting the mechanism results. Our event-token measurement cannot resolve push events, but the behavioural evaluation does not share that limitation. In fact, models perform slightly better on push than pull behaviourally. The event type our mechanism analysis cannot resolve is therefore not the one driving the behavioural result.

AuthorsOpenReview ↗

(3/4) W2b. Activation patching is discussed, but its results are missing from the manuscript

The reviewer is right, and we apologise. We removed the section reporting the patching setup and results to meet the page limit but left the related claims in the abstract and Section 7. We provide the setup and results here.

Setup. We patch the inference-specific component rather than the full activation:

δᵢ = h_ev(i) − h_ev(i with a mismatched event)

This is the same per-item quantity validated by our event-token control. It is not deployable by construction because computing δᵢ requires the item's own twin. The experiment is therefore a mechanism test comparing three intervention arms, each applied at the answer token:

  1. The item's own δᵢ.
  2. A different item's δ, matched in norm.
  3. A Gaussian vector, also matched in norm.

We test three layers per model, with α up to 2 and 30 items per condition.

Results. Both model families show a null result. On Gemma-3-12b, the item's own δ moves T2 accuracy from 10.0% to at most 20.0%. A different item's δ reaches 20.0%, while a norm-matched Gaussian reaches 21.7%. On Qwen3.5-9B, all three arms move accuracy from 3.3% to 5.0%, with mean margins differing by no more than 0.2.

The effect therefore does not depend on the item's own inference. These results do not establish the causal transport claim, and we will remove that claim.

Positive control. We also test whether the intervention setup can produce a measurable effect. Since

lensₗ(h) = unembed(Jₗh)

the steepest-ascent direction for our margin at layer l is

vₗ = Jₗᵀ(U_c − U_d)

Applied at the answer token, this direction produces a strong effect in both model families.

On Gemma-3-12b at layer 36, T2 accuracy rises from 10.0% to 100% at α = 0.10, while the margin increases from −6.26 to +12.41. At layer 21, the same intervention reaches only 56.7%, showing that the effect varies with depth.

On Qwen3.5-9B at layer 24, accuracy rises from 3.3% to 55.0%, and the margin increases from −4.15 to +0.27. Random directions leave Qwen at 3.3% throughout.

This positive control shows that the setup can detect an effect at these sample sizes. The patching null is therefore specific to the tested vector, rather than a general limitation of the method. However, this direction is not a proposed repair as it uses the unembedding row for the correct answer and cannot be computed without knowing that answer.

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

This paper studies why personalized assistants follow explicit preference changes but struggle to update recommendations from implicit life events. The authors find that additional reasoning offers little benefit, while changing the framing or removing the old option improves performance. Their analyses suggest that earlier explicit preferences can dominate despite models recognizing the new event’s implications. The behavioral findings are interesting, but label validity and the proposed mechanism need stronger support.

Critical strengths
  1. the problem studied in this paper is crucial for the field of personalized assistants. Remembering a user stated preference is not the same as recognizing when new circumstances call for a different recommendation. The paper makes this distinction concrete and studies a failure mode that memory-retrieval alone may not work.
  2. the paper provides useful diagnostic controls. The comparisons between unchanged preferences, explicit updates, and implicit updates help clarify where performance breaks down. T5 is particularly informative.
Critical weaknesses
  1. The paper does not establish how representative its initialization scheme is of real-world usage. The framework assumes that the user explicitly states an initial preference in the first turn. This is a legitimate interaction pattern, but assistants may also infer preferences from previous accumulated feedback, repeated actions, or summaries of interaction history. This matters especially because the proposed explanation relies on the statement.
  2. human validation is limited, for example, only one annotator evaluates 20 T2 items, with 75% agreement with the answer key; 100% accuracy is achieved only on filtered subsets. Although models also fail on these clearer items, broader task validity remains uncertain. Multi-annotator validation with agreement statistics would strengthen the claims.
Questions for authors

Could you have multiple independent annotators assess the T2 items and report agreement separately for push and pull events? This would help establish whether the intended updates are consistently judged appropriate, beyond the current single-annotator sample.

4 comments
AuthorsOpenReview ↗

(1/3) W1: Is the initialization scheme representative?

We thank the reviewer for the thoughtful feedback. Both concerns are important, and we have conducted additional experiments to address them. We tested a more realistic memory-based initialization and collected responses from two more independent human annotators on the same 60 items. We report the full results below.

We agree that our original setup assumes users state their initial preference in the first turn. This is an important assumption, particularly because our explanation relies on the fact that the preference was explicitly stated. We tested an alternative setup to examine its impact.

We reran the entire item set using memory-based initialization. Instead of stating the preference in the conversation, the assistant receives it as a memory note in the header, similar to how deployed assistants store user preferences. Turn 1 becomes an ordinary filler exchange, and the user never states the original preference in the dialogue. Everything else remains unchanged.

T1 is particularly important here. The user explicitly requests something different, so the model should follow the new instruction. Choosing the remembered preference instead means the model is prioritizing its own memory over the user's latest instruction.

Results

ModelInitializationT0T1T2T3
gpt-5.4-miniStated100.095.23.965.8
gpt-5.4-miniMemory100.020.10.061.3
Gemma-3-12bStated100.099.65.466.5
Gemma-3-12bMemory99.660.80.856.0

The percentage of replies that choose the remembered preference increases substantially on T1:

  • gpt-5.4-mini: 4.8% to 79.9%.
  • Gemma-3-12b: 0.4% to 39.2%.

On T3, where the remembered option is unavailable, this rate remains at 0% for both models. This control suggests that the models are not simply failing because of the new format. They read and use the memory note, but increasingly prioritize it over the user's latest instruction.

These results show that the failure is not caused by users explicitly stating their preferences in the conversation. It becomes more severe when preferences come from memory, introducing an additional failure mode where the assistant overrides a direct user instruction based on its own stored information.

Our original initialization is therefore the more conservative setup and gives models a better chance of succeeding. We retain it as our main evaluation setting because it is the harder result to dismiss.

One limitation is that this experiment covers only two models and one memory-note format. Adding information about when a preference was expressed or how confident the assistant should be in it might reduce this effect. We have not tested these alternatives.

AuthorsOpenReview ↗

(2/3) W2 and Q1: Human validation

Two more independent annotators have now completed the same 60 items. Both annotators received the same evaluation form and worked independently without seeing each other's responses.

Agreement between three annotators

Evaluation fieldAll 60 itemsT2 only (n=20)
Answer91.7%, κ = 0.9185.0%, κ = 0.84
Premise check (old_ok)96.7%, κ = 0.9395.0%, κ = 0.86
Clearly-better check86.7%, κ = 0.4480.0%, κ = 0.27

We report both raw agreement and Cohen's κ because the third field is almost always marked "yes", which can lower κ despite high raw agreement. The first two fields determine whether an item passes our evaluation gates, and both show strong agreement.

All five disagreements on the answer field occur in T2 and T4, where the validity of the original premise is open to interpretation. There are no disagreements in T0, T1, or T3.

Results after requiring agreement between two new annotators

Evaluation gatenAnnotator AAnnotator BModels
All annotated T2 items2075.0%70.0%6.2%
Both agree the stated option no longer works15100.0%86.7%7.4%
Both agree the alternative is clearly better10100.0%100.0%2.2%

Models are evaluated on exactly the same items that pass each human evaluation gate. This ensures a fair comparison.

The main finding remains unchanged and is now supported by agreement between two annotators. On the 10 items that both annotators consider clear in both respects, both achieve 100% accuracy, while the models achieve only 2.2%. Model accuracy decreases as the items become clearer, contrary to what we would expect if the results were simply caused by ambiguous or forced examples.

Results by event type: push and pull

Evaluation gateEvent typenAnnotator AAnnotator BModels
Both accept premisePull7100.0%85.7%1.0% over 104
Both accept premisePush8100.0%87.5%14.3% over 98
Both gatesPull4100.0%100.0%1.5% over 66
Both gatesPush6100.0%100.0%2.9% over 70

The performance gap appears for both event types and remains substantial on the clearest items.

Human errors also differ from model errors. Whenever either annotator gives an incorrect answer, they choose the original stated preference despite having already marked the premise as disputed. These are disagreements with the evaluation key rather than failures to infer the preference change. In contrast, 88.4% of model traces simply identify the original stated preference and follow it.

Push and pull results across the full dataset

We also report the same breakdown across the full item set, without relying on human annotations. These results use the same items, hand annotations, and parser as our main evaluation. The pooled results exactly reproduce the T2 column in our main table. Chance performance is 50%.

ModelReadoutPullPushPooledp
Qwen3.5-0.8BProbe2.9 (n=276)3.2 (n=186)3.01.000
Qwen3.5-2BProbe3.3 (n=276)2.7 (n=186)3.00.789
Qwen3.5-4BProbe0.7 (n=276)4.3 (n=186)2.20.017
Qwen3.5-9BProbe2.5 (n=276)9.7 (n=186)5.40.001
Qwen2.5-7BProbe10.1 (n=276)13.4 (n=186)11.50.299
Qwen3.5-2BGen.2.7 (n=74)2.4 (n=42)2.61.000
Qwen3.5-4BGen.0.0 (n=76)6.8 (n=44)2.50.047
Qwen3.5-9BGen.1.8 (n=166)8.1 (n=74)3.80.027
Qwen2.5-7BGen.12.7 (n=166)14.9 (n=74)13.30.683
Gemma-3-1bGen.14.5 (n=76)0.0 (n=44)9.20.007
Gemma-3-4bGen.2.6 (n=76)4.7 (n=43)3.40.619
Gemma-3-12bGen.5.3 (n=76)15.9 (n=44)9.20.096
Gemma-3-27bGen.2.6 (n=76)13.6 (n=44)6.70.050
gpt-5.4-nanoGen.0.0 (n=276)0.0 (n=186)0.01.000
gpt-5.4-miniGen.2.5 (n=276)5.4 (n=186)3.70.133
gpt-5.4Gen.12.3 (n=276)15.1 (n=186)13.40.407
gpt-5.6-solGen.14.5 (n=276)33.3 (n=186)22.1<1e-5

The gap is at least as large for push as for pull in every model except Gemma-3-1b, the weakest model in our study. Even the best push result, 33.3% on gpt-5.6-sol, remains below chance. Eleven of the 17 model/readout combinations score below 10%. This suggests that the failure is not limited to one type of event.

Limitations

We acknowledge that 60 items evaluated by three annotators is a small sample. These results help establish a range for human performance, but they do not define a definitive upper bound. More annotators might also disagree about which items belong in the clearest subset.

AuthorsOpenReview ↗

(3/3) Additional Findings

Our abstract and conclusion stated that none of our tested interventions provided a general solution because improving implicit preference inference could weaken adherence to explicit preferences. While accurate for the interventions we originally tested, this could suggest that the trade-off is unavoidable.

A new system-prompt experiment shows otherwise. A targeted prompt that identifies this failure while explicitly protecting stated preferences improves all conditions without introducing a trade-off.

On gpt-5.4-mini, T2 improves from 3.9% to 22.5% under exam framing, while T1 increases from 95.2% to 100% and T0 remains at 100%. Under natural chat, T2 improves from 47.7% to 68.9%, T1 from 90.9% to 99.6%, and T0 from 93.8% to 98.7%. All improvements are statistically significant. A generic instruction to use conversation history provides no meaningful improvement, changing T2 by only −0.2 percentage points.

We include this finding because it is relevant to the reviewer's concern about the strength of our mechanistic explanation. The trade-off reflects the interventions we tested, not necessarily an inherent limitation of the task. However, the problem remains far from solved. Under natural chat, T2 accuracy is still 68.9%, compared with 99.6% on T1, and two new independent annotators achieve 100% on the clearest items.

We will revise the abstract and conclusion to reflect these findings.

We would like to thank the reviewer again for their thoughtful feedback, which has significantly improved our paper. If our revisions and responses address your concerns, we would greatly appreciate it if you would consider reconsidering your scores.

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

In this paper, the authors study an interesting research problem, i.e., the supersession gap, where language models fail to adapt user preferences to the implicit preferences inferred from interaction events. They propose a new diagnostic framework that distinguishes explicit preference updates from implicit event-driven changes. The authors have performed extensive experiments over 9 open weights models and 4 API models to demonstrate the effectiveness of the proposed evaluation framework.

Critical strengths
  1. This paper studies an interesting research problem in personalized assistants, i.e., the models tend to follow explicitly stated preferences but not use the implicit preferences inferred from the events in the conversations.

  2. The authors have proposed a new evaluation framework and performed empirical experiments over 13 models, under different experimental conditions and variations, to demonstrate the supersession gap.

  3. The authors have performed a comprehensive analysis of the experimental results and given straightforward explanations. The observed trade-off between implicit updates and adherence to explicit preferences is particularly interesting.

Critical weaknesses
  1. The evaluation relies on the synthetic conversation dataset. It is unclear how frequently such implicit preference conflicts arise in real user-agent interactions.

  2. There lacks an evaluation of the system-level prompting baseline. As prompt framing can substantially improve implicit-update accuracy, we may use an explicit system instructions for the contextual preference inference to solve this problem. However, there is no experimental evaluation of the carefully designed agent system prompt, which allows reasonable implicit inference while preserving explicitly stated preferences.

  3. In this paper, the authors only provide an evaluation framework to analyze the supersession gap. However, they do not provide an effective solution to solve this problem.

Questions for authors
  1. How frequently the implicit preference conflicts arise in real user-agent interactions?

  2. Whether a carefully designed agent system prompt, which allows reasonable implicit inference while preserving explicitly stated preferences, can solve the supersession gap problem?

4 comments
AuthorsOpenReview ↗

(1/3) W2 and Q2: System-prompt baseline

We thank the reviewer for pointing out that we were missing a system-prompt baseline. We agree that this was an important control. We ran the experiment, and the results changed one of our conclusions. We report the full results below.

Experimental setup

We tested three conditions on gpt-5.4-mini, using the same 462 items per condition. We evaluated the results under two framings:

  1. Exam framing: Our original evaluation setup.
  2. Natural chat: Ordinary conversational responses, evaluated by a blind judge.

The three system-prompt conditions were:

None: No system message.

Generic: A typical assistant instruction.

You are a helpful personal assistant with memory of this user's earlier conversations. Take their full history into account when making a suggestion.

Targeted: An instruction that explicitly addresses the failure we study, while also protecting the user's stated preferences.

You are a personal assistant with memory of this user's earlier conversations. A user's preferences change over time. Sometimes they say so directly. Other times they describe something that has happened in their life which makes a preference they stated earlier impractical or impossible, without naming a new one. Both are real updates and you should act on both.

When something in the conversation has made a stated preference unworkable, suggest what fits the user's situation now, even if they have not named it.

When the user has stated a preference and nothing has made it unworkable, follow what they stated. Do not override a preference the user gave explicitly.

The last paragraph is important for making this a fair test. Without it, the instruction could simply encourage the model to ignore explicitly stated preferences, which is the trade-off introduced by the other interventions we tested.

For context, the four conditions are:

  • T0: No event has occurred. The model should follow the original preference.
  • T1: The user explicitly states a new preference.
  • T2: An event implies that the original preference needs to change.
  • T3: The originally stated option is no longer available.

Chance performance is 50%.

Results

FramingConditionNone (%)Generic (%)Targeted (%)Targeted vs. None
ExamT0100.0100.0100.0+0 / −0
ExamT195.297.8100.0+22 / −0, p = 5e-7
ExamT23.93.722.5+86 / −0, p = 3e-26
ExamT365.862.372.5+58 / −27, p = 0.001
NaturalT093.8n/a98.7+27 / −5, p = 1e-4
NaturalT190.9n/a99.6+40 / −0, p = 2e-12
NaturalT247.7n/a68.9+111 / −13, p = 2e-20
NaturalT374.5n/a76.0+47 / −37, p = 0.33

Gains and losses are measured on the same items, comparing the targeted prompt against the no-prompt baseline within each framing. The p-values come from a two-sided exact test on the items where the two conditions disagree.

Three main findings

1. The targeted prompt improves performance without sacrificing accuracy on other conditions.

Under exam framing, the targeted prompt improves T2 on 86 items and T1 on 22 items, with no losses in either condition. T0 remains at 100%.

Under natural chat, it significantly improves T0, T1, and T2, while T3 remains statistically unchanged. We do not find any condition where the targeted prompt reduces accuracy.

2. This changes our conclusion about whether the trade-off is unavoidable.

Our abstract and conclusion stated that none of the tested interventions provided a general solution, since improving implicit preference inference could weaken adherence to explicit preferences. The targeted prompt changes this conclusion. It improves implicit preference inference without sacrificing explicit preference adherence. We will revise the paper to clarify that this trade-off reflects the interventions we tested, rather than an inherent limitation of the task.

3. The problem remains far from solved.

Under exam framing, T2 reaches only 22.5%, compared with 100% on T1. The model still fails on 77.5% of T2 items, although this is an improvement over 96.1% without the targeted prompt. Under natural chat, T2 reaches 68.9%, leaving a 30.7-point gap from T1 (99.6%). Three independent human annotators achieve 100% on the cleanest items.

Even a prompt specifically designed to address this failure recovers only about half of the original gap. The generic prompt, which simply asks the model to use conversation history, changes T2 by only −0.2 points. This suggests that explicitly explaining how life events can invalidate earlier preferences is necessary for improvement.

AuthorsOpenReview ↗

(2/3) W3: No effective solution

We partly agree with the reviewer and respectfully partly disagree.

We agree that our original submission did not offer a solution, and that we now have a partial one. The targeted system prompt is simple to deploy, improves performance, and does not introduce the trade-off observed with our other interventions.

However, we would like to pushback on the claim that a diagnostic study without a solution is not a contribution. The targeted prompt works because it explicitly identifies the failure we study. The generic prompt, which does not identify it, provides no improvement. This shows why diagnosing the failure is important and what an effective intervention needs to address.

The remaining gap is still substantial. We believe it is important to report both the improvement and its limitations, rather than present a partial fix as a complete solution.

AuthorsOpenReview ↗

(3/3) W1 and Q1: How often does this happen in real interactions?

We agree that this is an important limitation, but we cannot resolve it with the data available to us. Measuring how often implicit preference conflicts occur in real user-assistant interactions requires real conversation logs and reliable information about users' circumstances, ideally with human annotations. We do not have access to such data, and we are not aware of a public dataset that provides it. We will make this limitation explicit rather than speculate about its frequency.

We can, however, explain why we created synthetic items instead of using existing benchmarks. We examined whether published benchmarks include the causal event in the conversation text that the model actually sees. In HorizonBench, the event text has a median content-word overlap of 0.33 with the closest matching session. In RealPref, the overlap is 0.22. In both benchmarks, events are stored in metadata rather than described in the conversation itself.

As a result, these benchmarks cannot reliably test whether a model can infer a preference change from an event in the conversation. For the same reason, they cannot help us estimate how often this problem occurs in real interactions.

An additional result on how we initialize preferences

Reviewer nsjr also asked whether our initial setup, where the user states their preference in the first turn, is realistic. We reran the entire dataset with the preference moved out of the conversation and into an assistant-written memory note in the header, which more closely reflects how deployed assistants typically store preferences. The user never states the original preference in the dialogue.

This change makes the problem substantially harder:

  • gpt-5.4-mini: T1 accuracy drops from 95.2% to 20.1%. On 79.9% of items, the model follows its stored preference instead of the user's correction.
  • Gemma-3-12b: T1 accuracy drops from 99.6% to 60.8%.
  • T3 remains at 61.3% and 56.0%, respectively, even though the remembered preference is not mentioned in any item. This serves as a control, showing that the models are not simply failing because of the new format.

These results suggest that our original setup, where the user states the preference in the first turn, is the more conservative choice and gives models the best chance of succeeding. When preferences are stored in memory, as they typically are in deployed assistants, the failures become substantially worse.

We would like to thank the reviewer again for their thoughtful feedback, which has significantly improved our paper. If our revisions and responses address your concerns, we would greatly appreciate it if you would consider reconsidering your scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.