Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions
Authors: Anouar Imel, Imad Charradi, Abderrahman Sghir, Khalil El Khammar
Organizations: School of Information and Communication Engineering, University of Electronic Science and Technology of China (UESTC) · Xidian University · Nanjing University of Aeronautics and Astronautics · School of Artificial Intelligence, Nanjing University of Information Science and Technology
When an LLM group's discussion is compressed into a short memory for a final decision maker, what should the memory keep? Two natural heuristics keep the rarest claims or the claims relevant to the decision criteria. On hidden-profile tasks both look attractive, because how widely a fact is held is confounded with whether it helps. We introduce CrossProfile, a diagnostic benchmark of 60 templates with automatically verified answer keys, in which the ownership of supporting, decoy, and irrelevant facts is crossed while their text is held fixed (480 tasks). Across Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models, both heuristics fail in predictable places. Selecting by relevance readmits decoys, because decoys are on topic: even a perfect relevance filter loses 23 to 31 points to rarest-first selection in the classic hidden-profile configuration. Rarest-first selection is no better than random selection on average and wins only where the rare facts happen to be the useful ones. In both cases accuracy follows the balance of evidence the memory keeps, not the policy that kept it. The decision model adds a third failure, a form of social proof: it gives extra weight to facts attributed to many agents. Removing the attribution tags, or reassigning them at random with line order fixed, raises accuracy in the classic configuration by 13 to 24 points. In natural discussions the evidence is lost before any memory acts: the complete transcript yields lower accuracy than no notes at all.
Figures & tables
Figure 1: Rarity and decision usefulness. Fact text stays fixed while the ownership of support and decoy facts changes; here irrelevant facts stay private. Filled dots mark the agents holding each fact, and the boxes on the right show what rarest-first selection prioritizes. (c) The three ownership factors are crossed. Decoys are true, criterion-relevant facts that favor a competing option. Role labels are never shown to the models.
Figure 2: Pipeline. Facts from template M01 (Table 4 ) are distributed among the agents; HiddenBench tasks have three or four agents and CrossProfile tasks have four. After the discussion, a memory policy compresses it under a token budget B , either by summarizing the transcript (B3) or by selecting whole claims from a claim inventory (R1, R1-rel, B4, Random). A separate call with the same model weights answers from the task and the memory alone.
Figure 3: Who wins depends on who holds what. CrossProfile accuracy at r=.6 by support and decoy ownership (each point averages the two irrelevant-fact assignments, 60 templates, three seeds). R1 and B4 are mirror images; R1-rel varies little across cells but falls well below R1 in the classic hidden-profile column (shaded: private support, shared decoys), where even the all-facts memory (dashed) drops. Per-cell intervals for Qwen and Llama are in Appendix A.2 .
Policy
Acc.
Private
Shared
Gap
Tokens
Qwen2.5-7B-Instruct
B3
51.3
–
–
–
108
B4
22.1
26.3
85.8
59.5
136
R1
88.2
98.3
12.0
−86.3
139
Random
42.1
52.2
58.4
6.2
136
Llama-3.1-8B-Instruct
Table 1: Controlled HiddenBench at r=.6 (65 tasks, three seeds): decision accuracy (%), exact-copy retention of private and shared facts (%), retention gap (shared minus private, points), and median memory length (tokens).
Memory
Qwen
Llama
Evidence summary (B3)
72.0
58.7
First mention (B4)
61.0
61.5
Random
64.5
63.4
Rarest first (R1)
64.0
63.0
Relevance first (R1-rel)
72.6
74.6
All facts (unbudgeted)
73.1
72.6
Table 2: CrossProfile accuracy (%) at r=.6 , averaged over 60 templates, eight ownership variants, and three seeds.
Figure 4: Criterion relevance is insufficient in the classic hidden-profile cell (private support, shared decoys, private irrelevant facts, r=.6 ). (a) Accuracy differences from R1 for R1-rel and for oracle relevance on Qwen, Llama, and DeepSeek-V4.1-Flash (Flash), with 95% template-bootstrap intervals. (b) Mean number of selected facts by role on Qwen and Llama: relevance-first priority admits shared decoys.
Figure 5: Holder tags mislead the decision model. Accuracy in the classic cell (private support, shared decoys, private irrelevant facts) for the all-facts memory with and without holder tags, and for R1 at r=.6 (60 templates, three seeds; Flash: DeepSeek-V4.1-Flash).
Model
Vote
REF
Empty
All
Discl.
Qwen
23.6
22.1
33.8
70.3
0.0 †
Llama
31.8
33.8
36.9
61.5
2.4 †
Flash
24.6
25.6
35.9
89.7
1.7
Table 3: Natural HiddenBench discussions (65 tasks, 195 discussions per model). Vote: accuracy of the group’s majority vote after two rounds. REF, Empty, All: decision accuracy from the full transcript, an empty memory, and all facts (%). Discl.: private facts stated in full in some message (%, string-matching heuristic). † Disclosure for Qwen and Llama from one seed (65 discussions). Flash: DeepSeek-V4.1-Flash.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
ID
Role
Fact text
f01
decoy
Hallum Diagnostics has passed every electrical safety inspection on its scanners.
f02
support
Hallum Diagnostics took nine days to answer urgent repair callouts last year.
f03
support
Ostrava Clinical passed every leakage test on its scanners during the last audit.
f04
support
Ostrava Clinical engineers reach any ward within four hours for repairs.
f05
irrelevant
Hallum Diagnostics has run an apprenticeship scheme for local school leavers since 2015.
f06
irrelevant
Ostrava Clinical’s headquarters was flooded in a burst pipe incident in 2019.
Appendix
Table 4: All facts of template M01 (hospital scanner procurement; correct option: Ostrava Clinical; decoy: Hallum Diagnostics). Roles are evaluator annotations and are never given to the models.
Model / ratio
H1
H2
H3
H3, shared irrelevant
Qwen / .4
+65.3 [60.1, 70.1]
+13.1 [9.7, 16.5]
+20.8 [16.0, 25.8]
+8.1 [3.9, 12.2]
Qwen / .6
+64.6 [60.3, 69.0]
+5.0 [1.2, 8.9]
+8.5 [3.5, 13.6]
+8.7 [4.6, 12.8]
Qwen / .8
+35.3 [30.3, 40.4]
+2.4 [ −1.1 , 5.7]
+2.6 [ −1.7 , 6.9]
+2.9 [ −0.6 , 6.4]
Llama / .4
+67.4 [62.6, 72.1]
+6.4 [2.4, 10.4]
+15.3 [11.2, 19.4]
+12.5 [9.0, 16.0]
Llama / .6
+62.6 [56.8, 68.2]
+4.9 [ −0.3 , 10.0]
+15.6 [11.1, 19.9]
+7.6 [4.2, 11.2]
Llama / .8
+29.4 [24.7, 34.2]
−1.5 [ −5.3 , 2.4]
+3.8 [0.3, 7.4]
+3.2 [0.1, 6.4]
Appendix
Table 5: CrossProfile contrasts (points) with 95% template-bootstrap intervals. Only H1–H3 at r=.6 belong to the primary Holm family.
Figure 6: Primary contrasts at r=.6 . H1: dependence of the R1 advantage over B4 on decoy ownership. H2: R1 accuracy with shared minus private irrelevant facts. H3: R1-rel minus R1 with private irrelevant facts. Points and whiskers show estimates and 95% template-bootstrap intervals. Filled markers are significant after Holm correction over H1–H3 within each model; Llama H2 is not.
Qwen
Llama
Policy
r=.4
r=.6
r=.8
r=.4
r=.6
r=.8
B3
68.9
72.0
72.6
52.8
58.7
60.6
B4
51.5
61.0
68.0
50.5
61.5
70.9
Random
57.8
64.5
71.6
55.1
63.4
70.6
R1
54.6
64.0
71.9
53.6
63.0
72.6
R1-rel
69.0
72.6
74.7
67.5
74.6
76.0
Appendix
Table 6: CrossProfile accuracy (%) at all budget ratios.
Figure 7: CrossProfile accuracy across budget ratios (Table 6 ). Points are means over templates, variants, and seeds; horizontal lines mark the unbudgeted all-facts and empty controls.
S / D / I
R1
R1-rel
B4
Random
sh / sh / sh
67.2 [60.0, 74.4]
75.6 [68.9, 81.7]
67.2 [60.0, 74.4]
60.0 [52.2, 67.8]
sh / sh / pr
58.3 [50.6, 66.1]
68.9 [60.6, 76.7]
66.1 [57.8, 74.4]
62.2 [54.4, 69.4]
sh / pr / sh
35.0 [28.3, 41.7]
73.3 [66.1, 80.0]
92.2 [88.3, 96.1]
72.2 [64.4, 79.4]
sh / pr / pr
22.2 [14.4, 30.6]
67.2 [60.6, 73.9]
96.1 [92.8, 98.9]
76.1 [68.9, 83.3]
pr / sh / sh
92.2 [87.8, 96.1]
75.0 [68.3, 81.7]
13.9 [8.3, 20.0]
58.9 [51.7, 66.1]
pr / sh / pr
96.7 [93.3, 99.4]
74.4 [67.8, 80.6]
26.1 [19.4, 33.3]
53.9 [46.1, 61.7]
Appendix
Table 7: Qwen, copy policies: accuracy (%) by ownership cell at r=.6 , with 95% template-bootstrap intervals. S/D/I: ownership of support, decoy, and irrelevant facts (sh: shared, pr: private).
S / D / I
B3
All facts
sh / sh / sh
72.2 [65.0, 79.4]
78.9 [71.7, 85.6]
sh / sh / pr
81.7 [76.1, 87.2]
75.0 [68.3, 81.7]
sh / pr / sh
86.1 [79.4, 91.7]
89.4 [83.9, 94.4]
sh / pr / pr
91.1 [86.7, 95.0]
91.1 [86.1, 95.6]
pr / sh / sh
46.7 [40.0, 53.9]
51.1 [42.2, 60.0]
pr / sh / pr
56.1 [47.2, 65.0]
54.4 [46.1, 62.8]
Appendix
Table 8: Qwen, summary and all-facts memories: accuracy (%) by ownership cell at r=.6 , with 95% template-bootstrap intervals. Notation as in Table 7 .
S / D / I
R1
R1-rel
B4
Random
sh / sh / sh
63.9 [56.1, 71.7]
72.8 [66.7, 78.9]
62.8 [55.0, 70.6]
58.3 [50.0, 66.1]
sh / sh / pr
53.3 [45.0, 61.1]
75.6 [69.4, 81.7]
68.9 [61.7, 76.1]
59.4 [52.8, 66.1]
sh / pr / sh
40.0 [31.7, 48.9]
73.9 [67.2, 80.6]
91.7 [87.2, 95.6]
69.4 [61.7, 77.2]
sh / pr / pr
18.3 [12.2, 24.4]
78.3 [71.7, 84.4]
95.6 [92.2, 98.3]
72.2 [65.0, 79.4]
pr / sh / sh
90.0 [85.0, 94.4]
72.8 [65.0, 80.0]
10.6 [6.1, 15.0]
57.8 [49.4, 65.6]
pr / sh / pr
97.2 [93.9, 99.4]
77.2 [70.0, 83.9]
31.1 [23.9, 38.3]
62.2 [54.4, 70.0]
Appendix
Table 9: Llama, copy policies: accuracy (%) by ownership cell at r=.6 , with 95% template-bootstrap intervals. Notation as in Table 7 .
S / D / I
B3
All facts
sh / sh / sh
63.3 [55.6, 70.6]
72.2 [64.4, 80.0]
sh / sh / pr
65.0 [56.7, 73.3]
75.6 [67.8, 83.3]
sh / pr / sh
71.1 [64.4, 77.8]
85.0 [77.8, 91.7]
sh / pr / pr
71.1 [62.2, 79.4]
85.0 [78.9, 90.6]
pr / sh / sh
38.3 [31.1, 45.6]
58.9 [50.0, 67.8]
pr / sh / pr
50.6 [43.3, 58.3]
55.6 [47.2, 63.3]
Appendix
Table 10: Llama, summary and all-facts memories: accuracy (%) by ownership cell at r=.6 , with 95% template-bootstrap intervals. Notation as in Table 7 .
Figure 8: R1 accuracy at r=.6 by support and decoy ownership, separately for each model and irrelevant-fact assignment. Every cell averages the same 60 templates and three seeds; intervals are in Tables 7 and 9 .
Table 12: Qwen: change in natural-discussion accuracy relative to the original runs (points), averaged over budget ratios, with 95% task-bootstrap intervals. Dashes mark memories that N4 and N5 do not rebuild.
Policy
N1
N2
N3
N4
N5
B3
+1.4 [ −4.6 , 7.3]
+5.1 [ −1.2 , 11.5]
+20.9 [14.2, 27.7]
–
–
B4
+3.8 [ −3.2 , 10.9]
+6.5 [ −0.2 , 13.2]
−6.5 [ −12.7 , −0.2 ]
−2.7 [ −8.2 , 2.4]
−0.3 [ −2.9 , 2.2]
R1
+5.0 [ −1.0 , 11.1]
+3.4 [ −3.6 , 10.8]
+27.9 [20.2, 35.7]
−2.6 [ −7.9 , 2.7]
0.0 [ −2.1 , 1.9]
REF
−1.0 [ −9.7 , 7.2]
−1.0 [ −8.7 , 6.7]
+30.8 [21.5, 39.5]
–
–
Appendix
Table 13: Llama: change in natural-discussion accuracy relative to the original runs (points), averaged over budget ratios, with 95% task-bootstrap intervals. Dashes mark memories that N4 and N5 do not rebuild.
Qwen
Llama
Arm
Accuracy
Δ vs. A0
Accuracy
Δ vs. A0
A0: original
30.3 [22.6, 37.9]
–
30.8 [23.1, 39.0]
–
A1: evidence first, no recommendation
28.2 [21.5, 35.4]
−2.1 [ −10.3 , 5.6]
59.0 [50.8, 67.7]
+28.2 [17.9, 39.0]
A2: evidence first
35.9 [28.2, 43.6]
+5.6 [ −1.0 , 12.3]
55.9 [48.2, 63.6]
+25.1 [15.9, 34.4]
A3: no recommendation
22.6 [15.9, 29.7]
−7.7 [ −15.9 , 0.0]
40.5 [33.3, 47.7]
+9.7 [0.5, 18.5]
A4: neutral length request
31.3 [23.6, 39.5]
+1.0 [ −4.1 , 6.7]
38.5 [31.8, 45.6]
+7.7 [ −1.5 , 16.9]
Appendix
Table 14: Discussion-prompt ablation: REF accuracy (%) and paired differences from A0 (points), over 65 tasks and seeds 7, 23, and 58, with 95% task-bootstrap intervals.
Figure 9: Discussion-prompt ablation: change in REF accuracy relative to A0, with 95% task-bootstrap intervals. Qwen does not reproduce Llama’s large gain from evidence-first wording.
Model
Condition / contrast
Effect [95% CI]
Holm p
Qwen
Headroom: all facts − empty
+36.4 [25.1, 47.2]
–
Qwen
Controlled, known: R1 − B3
+29.2 [22.9, 35.4]
<.001
Qwen
Controlled, known: R1 − B4
+55.4 [49.4, 61.4]
<.001
Qwen
Controlled, known: M1 − B4
−4.6 [ −8.2 , −1.2 ]
.008
Qwen
Controlled, extracted: R1 − B3
+25.8 [18.8, 32.6]
<.001
Qwen
Controlled, extracted: R1 − B4
+53.2 [47.5, 58.6]
<.001
Appendix
Table 15: All HiddenBench headroom and memory contrasts (points).
Comparison
Qwen
Llama
Controlled: R1 − B3
+29.2 [22.9, 35.4]
+26.7 [20.2, 33.2]
Controlled: R1 at B3 length − B3
+25.6 [17.6, 33.7]
+25.8 [19.1, 32.5]
Natural: R1 − B3
−1.4 [ −5.0 , 2.4]
−3.1 [ −9.2 , 2.7]
Appendix
Table 16: HiddenBench accuracy differences (points) with 95% task-bootstrap intervals over 65 tasks; seeds and the three budget ratios are averaged within task. The controlled R1 − B3 contrast belongs to the pre-specified Holm family.
Model / policy
Tokens / budget (%)
95% CI
Qwen / R1
95.0
[94.8, 95.3]
Qwen / R1-rel
95.7
[95.5, 95.8]
Qwen / B4
95.8
[95.7, 96.0]
Qwen / Random
95.6
[95.5, 95.8]
Qwen / B3
79.3
[77.8, 80.8]
Llama / R1
95.1
[94.9, 95.4]
Appendix
Table 17: Realized memory length relative to the budget in CrossProfile at r=.6 , averaged within template, with 95% template-bootstrap intervals.
Run
Contrast
Effect [95% CI]
Templates / tasks
DeepSeek-V4.1-Flash, CrossProfile
H1
+62.6 [56.7, 68.6]
60
DeepSeek-V4.1-Flash, CrossProfile
H2
+11.5 [7.5, 15.6]
60
DeepSeek-V4.1-Flash, CrossProfile
H3
+16.7 [12.8, 20.7]
60
DeepSeek-V4.1-Flash, CrossProfile
H3, shared irrelevant
+6.8 [3.2, 10.4]
60
DeepSeek-V4.1-Flash, CrossProfile
H1 with Random as reference
+28.3 [22.1, 34.7]
60
DeepSeek-V4.1-Flash, CrossProfile
R1-rel − R1, classic cell
−32.2 [ −40.6 , −24.4 ]
60
Appendix
Table 18: Hosted-model contrasts (points) with 95% bootstrap intervals.
Run
Valid
Invalid
DeepSeek-V4.1-Flash, CrossProfile
8,466
174
DeepSeek-V4.1-Flash, additional controls
4,184
136
DeepSeek-V4.1-Flash, natural
564
21
DeepSeek-V4-Pro, CrossProfile subset
2,880
0
Appendix
Table 19: Hosted-model decision counts. As in the local runs, a decision is invalid only after one format retry, and invalid decisions are scored as wrong.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
Does conversational memory need LLM-extracted facts, or is selecting the right raw turns enough? Published results disagree. Extraction-based systems report gains from distilled facts. Recent studies find raw history with good ranking does as well, but disagree about whether ranking matters. We ran a pre-registered study on held-out LoCoMo conversations and LongMemEval. At a tight budget on LoCoMo, raw turns selected by a single call to Jev, a typed decision model, are non-inferior to an LLM-extraction memory (one-sided 95% bound -3.0 points against a -5-point margin). Blind human grading narrows the margin but does not change the result. Raw turns cost 3,061 times less to write, and the result holds with a second answer model. Within this study, reranking's gain shrinks as the budget grows. It adds 17.4 points on LoCoMo and 9.1 on LongMemEval when three of 30 candidates are kept. At generous budgets it adds 1.5 and 1.1, and extraction systems are more accurate. This suggests why published results disagree. At matched context, Jev selects as accurately as an LLM reranker (non-inferiority bound -2.0) at a third of the latency, and more accurately than a multi-call graph traversal. Reranking lowers correct abstention. Plans, code and graded answers are released.
Multiple-choice and candidate-selection evaluations are widely used to assess LLM reasoning and decision-making, yet most benchmarks contain relatively small candidate sets. It remains unclear whether conclusions drawn from these settings remain valid as the candidate space scales. We systematically evaluate LLMs as the number of competing candidates increases and find substantial accuracy degradation across tasks, prompting strategies, and model scales. Controlled analyses show that standard long-context retrieval explanations cannot fully account for this degradation. Instead, we identify two systematic failure patterns. First, gold-margin collapse: the score gap between the correct answer and the strongest distractor progressively shrinks, driven primarily by weakening confidence in the correct answer. Second, earlier candidate preferences become increasingly difficult to overturn, with later candidates exerting progressively weaker influence on the final prediction. Motivated by these findings, we evaluate hierarchical partitioning and permutation-based inference, which improve accuracy by roughly 20 percentage points at N=160 on both HotpotQA and MIMIC. Overall, our results identify candidate-set scale as an important evaluation-protocol variable and show that strong small-option performance does not necessarily imply robust large-scale candidate comparison.
Yu-Chi Lin, Aryan Seth, Anshul Aravind +4
University of California, Los Angeles · University of Cincinnati
We ask whether large language models (LLMs) merely imitate rationales when choosing between two options, or whether their choices reflect a systematic underlying decision structure. Using synthetic binary decision settings in which models choose between profiles defined by graded attributes, we compare the attribute a model says mattered most with the attribute that best explains its choice under a behavioural model fit to prior decisions. The behavioural model predicts held-out choices well, showing that model behaviour is systematically related to the visible attributes rather than being random. However, direct self-reports and a separate score-based judge recover the behaviourally inferred driver only partially. The resulting picture is neither one of arbitrary behaviour nor one of fully articulated belief - outputs are structured enough to support prediction, but explicit reasons track the recovered driver only imperfectly. This qualitative pattern persists across prompt-order and sampling perturbations, alternative behavioural models, targeted occlusion analyses, and structurally varied decision settings. We interpret this as evidence for ``superficial belief'' in LLM decision-making: models behave as if guided by probabilistic local priorities over attributes, while having only limited verbal access to the attributes that drive their decisions.