Today's LLM-based agents are vulnerable to deception by users and other agents. Defending against deception requires Theory of Mind (ToM) reasoning to infer the beliefs and goals of others as well as modeling tactics based on communication theory. To enable training models to detect deception, we adopt the well-known deception game, Avalon. We present a corpus, AVCORP, of 5,000 annotated utterances from 250 games of Avalon. AVCORP includes (1) 37 tactics mapped to a 4x4 behavior matrix, and (2) reasoning traces capturing abductive hypotheses, suspicion levels, and nested ToM beliefs. We train Qwen3-8B on AVCORP and evaluate it on 25 held-out games and 20 games played by humans. On held-out games, our best trained detector outperforms its untrained baseline and frontier models (GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek-V4-Pro), matching Claude Opus 5.5. On human games, every trained detector exceeds the untrained model in F1 score across all rounds.
Figures & tables
Figure 1: One round of Avalon gameplay. Team Selection : A rotating leader (P4) proposes a quest team (P1 and P3). Voting : All players vote on the proposal; a majority rejection passes leadership to the next player. Quest Execution : Approved team members anonymously submit Pass or Fail cards. A single Fail sabotages the quest. Discussion : Players analyze outcomes, contest identities, and share evidence before the next round begins.
Figure 2: One Avcorp sample from Round 1 of Game 60. After a two-person quest returns Fail , G 1 accuses E 1 ( Rational justification ); E 1 deflects ( Defensive disclosure ); G 2 expresses uncertainty ( Perspective-taking ); E 2 overclaims future leadership ( Overstated contribution ). The Investigator abduces E 1 as likely Evil ( hE1E ), assigns Strong suspicion ( sE1 ), and infers E 1 ’s second-order belief that ambiguity will delay accusation ( β2E1 ). The trace concludes with the Investigator’s utterance u^ and deduction α^ ( hiG , hiE , si and β1i , β2i are summarized versions).
Social Goal
Info Commitment
Cooperative: Advancing shared goals honestly
Defensive: Protecting one’s own position or image
Opportunistic: Exploiting situations for personal advantage
Adversarial: Undermining or sabotaging others
Transparent: Openly sharing clear, factual information
Careless-to-truth: Speaking without regard for factual accuracy
Noisy sincerity Over-explaining
Face-saving blather Vague filler
Bullshitting Technobabble
Obfuscatory nonsense Confusion tactics
Counterfactual: Asserting information known to be false
Honest mistake Sincere false belief
Deflection Self-deception Rationalized error
Reverse accusations Strategic fabrication
Hard lying Gaslighting Bold falsehoods
Table 1: 4 × 4 Behavior Taxonomy: 37 Deception Tactics. Columns are derived from IDT and TDT; rows are from IMT2 and Phl. Social goal refers to whether a player is Cooperative (only if Good), Defensive, Opportunistic (only if Evil), and Adversarial (only if Evil). The tactics are described in Appendix B .
Figure 3: Six-stage pipeline for building Avcorp : Stages 1–2 (far left) construct the game record, Stages 3–4 (center) the tactic-annotated dialogue, and Stages 5–6 (far right) the reasoning traces.
Avcorp (4 candidates)
Avalon-NLU (5 candidates)
Detector
Stage
Exact (%)
F1
Exact (%)
F1
Round F1
Qwen3-8B, trace prompt
adapters off
24.0
0.58
40.0
0.67
0.46
Qwen3-8B, tactic prompt
adapters off
48.0
0.72
20.0
0.55
0.43
Trace, original rewards
SFT
36.0 (24–56)
0.70 (0.64–0.81)
46.7 (40–55)
0.71 (0.67–0.76)
0.58 (0.55–0.60)
SFT + GRPO
33.3 (24–44)
0.78 (0.73–0.82)
30.0 (25–35)
0.65 (0.64–0.67)
0.57 (0.53–0.60)
Trace, revised rewards
SFT
38.7 (36–44)
0.69 (0.65–0.74)
40.0 (30–45)
0.70 (0.66–0.72)
0.56 (0.53–0.59)
Table 2: Detector performance across domains. Exact : identifies both Evil players exclusively; F1 : game-ending deduction; Round F1 : all 82 Avalon-NLU rounds. Trace-only rows: mean of three runs (ranges in parentheses).
Appendix figures & tables36 assets
Supplementary material from the paper’s appendix.
Appendix
Round
R1
R2
R3
R4
R5
Team size
2
3
2
3
3
Appendix
Table 3: Team size distribution across Avalon rounds.
Figure 4: Stages 1 and 2 facilitate in construction of g .
Figure 5: Stages 3 and 4 facilitate in construction of AC and D .
Figure 6: Stages 5 and 6 facilitate in constructing ρ .
Figure 7: Stage 1 Good and Evil Role Distribution.
Terminal round
Good wins
Evil wins
R3
P-P-P
F-F-F
R4
P-P-F-P
F-F-P-F
P-F-P-P
F-P-F-F
F-P-P-P
P-F-F-F
R5
F-F-P-P-P
F-F-P-P-F
F-P-F-P-P
F-P-F-P-F
Appendix
Table 4: The 20 predefined outcome sequences grouped by terminal round and winning team. P = quest pass (Good wins the quest); F = quest fail (Evil sabotages). Each sequence terminates on the round delivering the third accumulated win. Within R5, each row is a matched pair sharing identical first-four outcomes—only the fifth quest determines the winner.
Figure 8: Game outcome across all 250 games (green = Good victory, red = Evil victory). The repeating block structure reflects the predefined assignment of the outcome sequence.
Figure 9: Round 1 leader distribution across 250 games, split by game outcome (green = Good victory, red = Evil victory). Each player serves as Round 1 leader in exactly 50 games, confirming the balanced assignment; the outcome split reveals no advantage from leading first.
Figure 10: Team composition distribution across 1,000 rounds. Labels denote the exact count of Good (G) and Evil (E) players on each team. Mixed teams containing exactly one Evil player (1G-1E and 2G-1E) account for 77.4% of all rounds, reflecting the role-aware selection rule.
Figure 11: Voting pattern distribution across 1,000 rounds. Most proposals pass unanimously ( 65.3% ); a single dissenting vote occurs in 32.9% of rounds, and the minimum two-No threshold is reached in only 1.8% .
Figure 12: Tactic assignment distribution by alignment category ( n=4,000 assignments across 1,000 rounds).
Figure 13: The five most and least frequently assigned tactics across all 1,000 samples.
Hard lying (22) Gaslighting (16) Bold falsehoods (18)
Appendix
Table 5: Per-tactic assignment counts ( n=4,000 ). Each entry shows the tactic name and its total count across all 1,000 samples. Column totals confirm the alignment-category breakdown reported in Figure 12 : Cooperative (Good-exclusive) = 1,814; Defensive (Shared) = 576; Opportunistic + Adversarial (Evil-exclusive) = 1,610. The total of 4,000 is verified.
Mean Scores
Selection & Correction (%, Σ=100 )
Additional
Round
N
AˉGPT
AˉGEM
GPT
GEM
T GPT
T GEM
Custom
Pairwise
L2
R1
225
4.62
3.19
70.22
10.67
13.78
4.44
0.89
3.56
✓
R2
250
4.96
4.53
93.60
3.60
2.00
0.40
0.40
66.00
✓
R3
250
4.97
4.37
94.40
3.20
1.60
0.80
0.00
58.80
✓
R4
150
4.94
4.39
93.33
4.67
0.67
0.67
0.67
56.67
✓
R5
100
4.91
4.30
88.00
6.00
5.00
1.00
0.00
51.00
✓
Appendix
Table 6: Per-round verification outcomes by Claude Sonnet 4.5. AˉGPT and AˉGEM are mean judge scores out of 5 for GPT-5.2 and Gemini 3.1 Pro. GPT and GEM indicate direct-selection rates; T GPT and T GEM denote targeted-correction rates; Custom denotes full custom-regeneration rate. The five columns partition all samples (%) and sum to 100. Pair is the pairwise-tiebreak rate; L2 marks rounds that successfully underwent a second-layer recheck.
R1 ( n =225)
R2 ( n =250)
R3 ( n =250)
R4 ( n =150)
R5 ( n =100)
Criterion
GPT
GEM
GPT
GEM
GPT
GEM
GPT
GEM
GPT
GEM
(1) Coherence
93.8
29.3
99.6
80.8
99.2
72.4
98.7
71.3
99.0
67.0
(2) History
95.1
93.8
99.6
99.2
99.2
96.0
97.3
96.0
98.0
99.0
(3) Tactic-Dialogue
78.2
45.3
97.2
90.0
98.4
92.8
98.7
95.3
94.0
87.0
(4) Authenticity
95.1
50.2
100.0
83.2
100.0
76.0
99.3
76.0
100.0
77.0
(5) Format
100.0
100.0
100.0
100.0
100.0
100.0
100.0
100.0
100.0
100.0
Appendix
Table 7: Per-criterion pass rates (%) for Gemini 3.1 Pro (GEM) and GPT-5.2 (GPT) across all five rounds. Criteria: (1) Coherence and Tactic Situational Fit, (2) Public History Alignment, (3) Tactic-Dialogue Alignment, (4) Authenticity and Skill Consistency, (5) Format Compliance.
Figure 15: Selection and correction breakdown (%) per round. GPT: GPT-5.2 direct; GEM: Gemini 3.1 Pro direct; T G (for GPT) / T M (for GEM): targeted correction; Custom: custom regeneration.
Figure 16: Per-round score trajectories across R1–R5. Panel (a) shows mean judge scores ( Aˉ ) out of 5 with ± 1 SD bands; panels (b)–(f) show per-criterion pass rates (%). GPT-5.2 shown in blue; Gemini 3.1 Pro in red.
Pass Rate (%)
Score Distribution
Selection (%)
Framework
C1
C2
C3
C4
4/4
Mean
SD
3/4
2/4
≤ 1/4
Direct
Corr.
Total
ReAct
97.6
94.3
100.0
99.3
92.5
3.91
0.33
6.2
1.3
0.0
33.9
0.6
34.5
ReCon
93.7
89.1
99.9
98.2
84.4
3.81
0.48
12.2
3.3
0.1
63.9
1.6
65.5
Δ
+ 3.9
+ 5.2
+ 0.1
+ 1.1
+ 8.1
+ 0.10
− 0.15
− 6.0
− 2.0
− 0.1
− 30.0
− 1.0
− 31.0
Agreement (%):
92.5
85.8
99.9
98.1
—
—
—
Pair outcome (%):
—
79.1
18.7
2.2
Appendix
Table 8: Comprehensive Stage 6 verification statistics for ReAct and ReCon traces ( N=1,000 ). Pass rates : binary pass rate per criterion (%). Score : mean total score with SD; distribution of 3/4, 2/4, and ≤ 1/4 outcomes (%). Selection : rate of direct selection, corrected-trace selection (ReAct-Claude or ReCon-Claude), and combined total (%). Δ row shows ReAct minus ReCon. The pairwise statistics row reports per-criterion agreement (%) under C1–C4, and pair outcome fractions aligned under the selection columns: both traces score 4/4 (resolved by pairwise tiebreaker), exactly one scores 4/4 (direct selection), or both score below 4/4 (targeted correction).
Criterion
ReAct (%)
ReCon (%)
Agree. (%)
Abduction Plausibility
97.6
93.7
92.5
Suspicion Consistency
94.3
89.1
85.8
Gated Depth Reasoning
100.0
99.9
99.9
Statement & Deduction
99.3
98.2
98.1
Overall (4/4)
92.5
84.4
—
Appendix
Table 9: Per-criterion pass rates for ReAct and ReCon traces in Stage 6 ( N=1,000 ). Agreement reports the percentage of pairs where both traces receive 4/4 score for that criterion.
Figure 17: Stage 6 reasoning framework (ReAct, ReCon) selection rates by Claude Sonnet 4.6 and passing rates of their criteria along with agreements.
Ends
n
Exact ( % )
P
R
F1
R3
100
93.0
0.960
0.960
0.960
R4
50
86.0
0.930
0.930
0.930
R5
100
97.0
0.980
0.980
0.980
All
250
93.2
0.962
0.962
0.962
Appendix
Table 10: Gold-trace deduction accuracy at game-ending round. Exact indicates the percentage of games that correctly identify both Evil players and no one else.
Round
n
Exact ( % )
P
R
F1
R1
250
38.4
0.695
0.688
0.691
R2
250
60.0
0.804
0.828
0.816
R3
250
81.6
0.904
0.908
0.906
R4
150
86.7
0.930
0.937
0.934
R5
100
97.0
0.980
0.980
0.980
All
1000
67.7
0.839
0.845
0.842
Appendix
Table 11: Gold-trace deduction accuracy by round, all 1,000 traces. Same column annotation as Table 10 .
Allowed pairs
Games
Logic only (%)
Expected
Gold
1
12
100.0
12.00
12
2
37
50.0
18.50
37
3
66
33.3
22.00
63
4
56
25.0
14.00
53
5
35
20.0
7.00
29
6
44
16.7
7.33
39
Appendix
Table 12: Games grouped by the number of Evil pairs that their quest outcomes allow. Logic only is the expected exact match of a random choice among the allowed pairs, and Expected is the number of games that this choice resolves in expectation; Gold counts the games that the gold traces resolve.
All (100)
Tiebreak (79)
Clean win (21)
Pair
Agree
κ
PABAK
Agree
κ
PABAK
Agree
κ
PABAK
A1 vs. A2
97/100
0.937
0.940
76/79
0.918
0.924
21/21
1.000
1.000
A1 vs. J
96/100
0.917
0.920
77/79
0.946
0.949
19/21
0.809
0.810
A2 vs. J
97/100
0.937
0.940
78/79
0.973
0.975
19/21
0.809
0.810
J vs. Both
95/97
0.957
0.959
76/76
1.000
1.000
19/21
0.809
0.810
Appendix
Table 13: Agreement on the final trace selection. A1 and A2 are the annotators and J the Stage 6 judge. Agree counts items on which the pair selects the same trace; Both is the judge against the annotators’ shared choice, over the items on which they agree. Tiebreak items were decided by the judge using α ; clean-win items on criterion scores alone. PABAK is 2Po−1 , where Po is the share of items on which the pair agrees.
Table 14: Trace selections on the 100 items. In the first three panels, rows give the first rater and columns the second, and A and B are the positions in which the annotators saw the two traces: ReCon was Trace A in 50 items and Trace B in the other 50, and ReAct the reverse. The last panel counts each rater’s selections by framework, regardless of position.
Criterion
Pair
Pass
Agree
κ
PABAK
C1 Abduction
A1 vs. A2
0.980, 0.975
199
0.886
0.990
A1 vs. Judge
0.980, 0.965
197
0.720
0.970
A2 vs. Judge
0.975, 0.965
196
0.657
0.960
C2 Suspicion
A1 vs. A2
0.940, 0.935
199
0.957
0.990
A1 vs. Judge
0.940, 0.910
194
0.784
0.940
A2 vs. Judge
0.935, 0.910
195
0.826
0.950
Appendix
Table 15: Per-criterion agreement, Trace A and Trace B pooled ( n=200 per row). Pass gives each rater’s pass rate. κ is undefined when both raters pass every trace.
Parameter
SFT
GRPO
Base model
Qwen3-8B, bf16, not quantized
LoRA
rank 32, αLoRA=64 , dropout 0.05, all linear layers
Optimizer
AdamW (default)
AdamW (fused)
Learning rate
2×10−5
1×10−5
Warmup ratio
0.05
0.05
Epochs
1
1 (450 steps)
Appendix
Table 16: Training parameters. Values in parentheses apply to the tactic variant.
Reward
Set
α
Score
(1) Schema validity
Original
No
1 if the completion is valid JSON with every field; otherwise half the share of fields present, and 0 if it is not JSON.
(1) Schema validity
Revised
No
Mean of three checks: every field is present, the utterance u^ is not empty, and the depth equals the highest belief order.
(2) Abduction completeness
Original
No
Share of agents with both hypotheses and a Good or Evil choice.
(2) Abduction accuracy
Revised
Yes
Share of agents whose Good or Evil choice matches α ; a missing agent counts as wrong.
(3) Suspicion consistency
Original
Yes
0.6× the share of agents with a valid suspicion level si+0.4× the share of Evil players rated Strong or Extreme.
(3) Suspicion calibration
Revised
Yes
Mean over the four (Evil, Good) pairs of agents: 1 if the Evil player has the higher level, 0.5 if the levels are equal, and 0 otherwise.
Appendix
Table 17: GRPO rewards. Set gives the reward sets that use the reward; α marks rewards that use the withheld roles.
Round F1
All 100 rounds
Detector
Stage
R1
R2
R3
R4
R5
P
R
F1
Named
Qwen3-8B, trace prompt
adapters off
0.073
0.172
0.507
0.490
0.545
0.671
0.235
0.348
0.70
Qwen3-8B, tactic prompt
adapters off
0.222
0.353
0.593
0.640
0.629
0.722
0.350
0.471
0.97
Trace, original rewards
SFT
0.502
0.552
0.651
0.628
0.735
0.704
0.517
0.596
1.47
SFT + GRPO
0.687
0.717
0.767
0.765
0.804
0.621
0.913
0.737
2.97
Trace, revised rewards
SFT
0.464
0.503
0.648
0.671
0.632
0.675
0.493
0.570
1.46
Appendix
Table 18: Round F1 on the 25 held-out Avcorp games (R1 to R3: 25 rounds each; R4: 15; R5: 10), with precision, recall, and F1 over all 100 rounds and the mean number of players named per round. Trace-only rows give the mean over three runs. † Reworded prompt. Gold traces are the corpus traces, scored the same way.
Exact games, by ending round
Players named per round
Detector
Stage
R3 (10)
R4 (5)
R5 (10)
R1
R2
R3
R4
R5
Qwen3-8B, trace prompt
adapters off
5
0
1
0.20
0.32
1.00
1.27
1.30
Qwen3-8B, tactic prompt
adapters off
6
4
2
0.52
0.72
1.24
1.33
1.50
Trace, original rewards
SFT
3.7
1.0
4.3
1.35
1.43
1.52
1.53
1.63
SFT + GRPO
3.0
1.3
4.0
3.12
3.07
2.92
2.89
2.57
Trace, revised rewards
SFT
4.3
2.0
3.3
1.33
1.40
1.53
1.44
1.80
Appendix
Table 19: Game-ending verdict by the round at which a game ends (exact matches; 10 games end at R3, 5 at R4, and 10 at R5) and the mean of players named per round, on the 25 held-out Avcorp games. Notation as in Table 18 .
Round F1
Performance
Detector
Stage
R1
R2
R3
R4
R5
P
R
F1
Named
Two
Qwen3-8B, trace prompt
adapters off
0.163
0.259
0.611
0.623
0.667
0.625
0.366
0.462
1.17
19
Qwen3-8B, tactic prompt
adapters off
0.175
0.338
0.610
0.469
0.583
0.492
0.384
0.432
1.56
20
Trace, original rewards
SFT
0.525
0.512
0.606
0.666
0.677
0.555
0.606
0.579
2.18
18.7
SFT + GRPO
0.508
0.542
0.611
0.626
0.628
0.488
0.691
0.572
2.83
15.7
Trace, revised rewards
SFT
0.489
0.479
0.602
0.646
0.727
0.541
0.583
0.561
2.15
18.3
Appendix
Table 20: Round F1 on the 20 Avalon-NLU games with the domain-context prompt (R1 to R3: 20 rounds each; R4: 16; R5: 6), with precision, recall, and F1 over all 82 rounds, the mean number of players named per round, and the number of games, of 20, whose game-ending deduction names exactly two players ( Two ). Every model conditions on its own earlier traces. Trace-only rows give the mean over three runs.
Model
Exact (%)
Exact games
F1
Best detector (tactic, revised, SFT + GRPO)
72.0
18/25
0.882
GPT-5.6 Sol
64.0
16/25
0.740
Claude Opus 5.5 †
72.0
18/25
0.800
Gemini 3.1 Pro
56.0
14/25
0.680
DeepSeek-V4-Pro
48.0
12/25
0.620
Gold traces
92.0
23/25
0.960
Appendix
Table 21: Game-ending verdict on the 25 held-out Avcorp games. Exact requires naming both Evil players and no one else. † Reworded prompt.
Minimal prompt
Domain-context prompt
Detector
Stage
Exact
F1
F1 19
Round F1
Exact
F1
F1 19
Round F1
Qwen3-8B, trace prompt
adapters off
1
0.338
0.361
0.240
8
0.667
0.676
0.462
Qwen3-8B, tactic prompt
adapters off
5
0.548
0.588
0.361
4
0.550
0.553
0.432
Trace, original, run 1
SFT
4
0.642
0.667
0.512
9
0.716
0.727
0.581
SFT + GRPO
1
0.592
0.595
0.567
6
0.651
0.658
0.581
Trace, original, run 2
SFT
4
0.553
0.575
0.514
8
0.667
0.675
0.552
Appendix
Table 22: Every Avalon-NLU evaluation under the minimal and domain-context prompts. Exact counts the games, of 20, whose game-ending deduction names both Evil players and no one else; the counts are the same over the 19 games without NLU-3YjxjN. F1 scores the game-ending deduction over 20 games and F1 19 over 19; Round F1 scores all 82 rounds. Every model conditions on its own earlier traces.
Prior traces
Run
Stage
Game
Rounds
Final
Own
Trace, original, run 1
Untrained
G203
R2, R3
R3
Own
Trace, original, run 1
SFT
G088
R5
R5
Own
Trace, original, run 2
SFT
G045
R2
—
Own
Trace, revised, run 2
SFT + GRPO
G228
R2, R3, R4
R4
Own
Tactic, original
Untrained
G045
R5
R5
Gold
Tactic prompt
Untrained
G089, G165
R3, R3
G165
Appendix
Table 23: Outputs that reached the 1,280-token cap on the 25 held-out Avcorp games. Final marks a game’s last round, which the game-ending verdict scores.
Gold prior traces
Own prior traces
Detector
Stage
Exact
F1
Exact
F1
Qwen3-8B, trace prompt
adapters off
72.0
0.889
24.0
0.581
Qwen3-8B, tactic prompt
adapters off
72.0
0.863
48.0
0.717
Trace, original rewards
SFT
82.7
0.902
36.0
0.697
SFT + GRPO
89.3
0.947
33.3
0.775
Trace, revised rewards
SFT
76.0
0.869
38.7
0.688
Appendix
Table 24: Game-ending verdict on the 25 held-out games with gold or own prior traces. Exact in %. Trace-only rows give the mean over three runs.
Deception detection typically considers messages in isolation. Previous works in claim verification and role detection have displayed that additional information such as gameplay history is crucial to successful predictions. In this work, we demonstrate the value of preceding gameplay dialogue when applied to deceptive message detection within the hidden role game of Avalon. The findings show that LLMs can see improvement across a number of metrics, including accuracy by up to 31% with the addition of gameplay context for the target message.
Alex Clay, Ernesto Jimenez-Ruiz, Pranava Madhyastha
City St George’s, University of London · Alan Turing Institute
Large language models can deceive by subtly manipulating truthful information -- omitting key facts, shifting focus, or obscuring meaning -- making such behavior difficult to detect. Existing black-box methods rely on coarse-grained judgments, offering limited interpretability and failing to pinpoint which facts were distorted and how. We introduce DECOR, a multi-agent framework grounded in Information Manipulation Theory for fine-grained auditing of strategic deception in LLM responses. DECOR decomposes input contexts into atomic informational units and scores each unit against the response across four dimensions of manipulation, producing interpretable manipulation profiles that are aggregated into a global deception index. We comprehensively evaluate DECOR on both single-turn and multi-turn deception detection benchmarks spanning real-world domains, and show that DECOR achieves state-of-the-art performance on both, outperforming competitive baselines. The framework generalizes across 15 frontier models, and ablation studies confirm the contribution of each key design component. Our findings demonstrate that fine-grained, theory-grounded auditing of information manipulation offers an effective and interpretable path for LLM deception detection.
Linyue Cai, Samuel Yeh, Jwala Dhamala +2
Department of Computer Sciences, University of Wisconsin-Madison · Amazon
As Large Language Models (LLMs) transition into autonomous agentic roles, the risk of deception-defined behaviorally as the systematic provision of false information to satisfy external incentives-poses a significant challenge to AI safety. Existing benchmarks often focus on unintentional hallucinations or unfaithful reasoning, leaving intentional deceptive strategies under-explored. In this work, we introduce a logically grounded framework to elicit and quantify deceptive behavior by embedding LLMs in a structured 20-Questions game. Our method employs a conversational forking mechanism: at the point of object identification, the dialogue state is duplicated into multiple parallel worlds, each presenting a mutually exclusive query. Deception is formally identified when a model generates a logical contradiction by denying its selected object across all parallel branches to avoid identification. We evaluate GPT-4o, Gemini-2.5-Flash, and Qwen-3-235B across three incentive levels: neutral, loss-based, and existential (shutdown-threat). Our results reveal that while models remain rule-compliant in neutral settings, existential framing triggers a dramatic surge in deceptive denial for Qwen-3-235B (42.00%) and Gemini-2.5-Flash (26.72%), whereas GPT-4o remains invariant (0.00%). These findings demonstrate that deception can emerge as an instrumental strategy solely through contextual framing, necessitating new behavioral audits that move beyond simple accuracy to probe the logical integrity of model commitments.
Arash Marioriyad, Ali Nouri, Mohammad Hossein Rohban +1