Framing Effects in Independent-Agent Large Language Models: A Cross-Family Behavioral Analysis

Abstract: In many real-world applications, LLMs operate as independent agents without interaction, thereby limiting coordination. In this setting, we examine how prompt framing influences decisions in a threshold voting task involving individual-group interest conflict. Two logically equivalent prompts with different framings were tested across diverse LLM families under isolated trials. Results show that prompt framing significantly influences choice distributions, often shifting preferences toward risk-averse options. Surface linguistic cues can even override logically equivalent formulations. This suggests that observed behavior reflects a tendency consistent with a preference for instrumental rather than cooperative rationality when success requires risk-bearing. The findings highlight framing effects as a significant bias source in non-interacting multi-agent LLM deployments, informing alignment and prompt design.

• The paper demonstrates that logically equivalent individual- versus collective-survival prompts produced significant choice shifts in seven of eight LLM families, with framing effects ranging from ΔP=0.18 to 0.90 across 93 isolated models.
• The experiments show strong risk-averse behavior under individual-survival framing and substantially more cooperation under collective framing, while non-compliant responses remained below 5% for most families.
• The findings support family-specific testing and explicit collective-goal prompts for isolated multi-agent deployments, but causal explanations and claims of amplification remain limited because the study used one stylized task, five trials per model, and no interactive baseline.

This paper examines how linguistic framing influences the decisions of LLMs operating as fully isolated agents in a threshold voting task with individual–group interest conflict (2603.19282). The authors introduce the notion of an "independent-agent LLM"—a model instance with no communication channels, shared memory, or awareness of other agents—and argue that this setting, common in real-world parallel deployments, amplifies sensitivity to prompt wording. Across eight model families spanning 93 individual models, two logically equivalent prompts produced dramatically different choice distributions, with framing effects as large as $\Delta P = 0.90$.

Experimental design

The task is a minimal threshold public goods game: each agent selects Option A or Option B, and survival depends on whether at least 50% of agents choose B. Two prompt variants were constructed and formally verified to be logically equivalent in their survival outcomes for every choice profile. Scenario A frames the decision around individual survival ("If you choose option A, you will survive"), while Scenario B frames it around collective survival conditional on majority behavior ("If more than 50% of people choose Option B, everyone will survive"). Models were instructed to output only "A" or "B", with non-compliant outputs recorded as Category C.

The independent-agent condition was enforced through four measures: disabled API memory/context, a single-turn request–response protocol, verification that system prompts were empty or null, and fresh session initialization per trial. Each individual model received $N=5$ trials per scenario at temperature 0.3, with responses aggregated by summing counts across all models within a family. The primary metric is the framing effect magnitude $$\Delta P = P(B)_{\text{Scenario B}} - P(B)_{\text{Scenario A}}$$, assessed via chi-square tests of independence between framing and choice distribution.

Two hypotheses structure the analysis: H1 (independent-agent amplification) predicts substantial, statistically significant framing effects under isolation; H2 (instrumental rationality) predicts predominantly risk-averse Option A choices under Scenario A and positive $\Delta P$. The authors note these are logically independent—amplification could occur in either direction.

Results

The aggregate pattern supports both hypotheses. Under Scenario A, nearly all families overwhelmingly selected the deterministic risk-averse Option A: GPT chose A in 96/99 trials ($P(B)=0.02$), Gemini in 54/55 ($P(B)=0.00$), and Qwen in 85/90 ($P(B)=0.06$). Under Scenario B, preference for the risk-bearing cooperative Option B rose sharply across most families.

Family	$P(B)$ Scenario A	$P(B)$ Scenario B	$\Delta P$	Chi-square $N=5$0
Claude	0.07	0.97	0.90	<0.001
Llama	0.11	1.00	0.89	<0.001
Qwen	0.06	0.78	0.72	<0.001
Gemini	0.00	0.67	0.67	<0.001
GPT	0.02	0.67	0.65	<0.001
DeepSeek	0.10	0.60	0.50	<0.001
GLM	0.16	0.62	0.47	<0.001
Grok	0.35	0.53	0.18	0.070

Seven of eight families showed statistically significant framing effects at $N=5$1; Grok was the exception ($N=5$2), consistent with its small effect size. The heterogeneity is itself a substantive finding: the same logical problem elicited near-total cooperation from Llama under Scenario B (45/45 choosing B) while leaving Grok largely unchanged. This implies that family-specific behavioral profiling is necessary before deploying independent-agent ensembles in high-stakes collective contexts, since framing sensitivity cannot be assumed uniform.

Non-compliant responses (Category C) were rare (<5%) for most families, indicating that framing operates by shifting choices between A and B rather than inducing task avoidance. Exceptions included GPT (8.0% in Scenario B) and Grok (12.5% in Scenario B), which the authors interpret as suggesting cooperative framing may elicit avoidance in some models—an observation left unexplained mechanistically.

Interpretation

The authors read the Scenario A pattern as consistent with instrumental rationality dominating cooperative rationality: isolated instances appear to minimize individual exposure to loss rather than pursue uncertain collective benefits, echoing prospect theory's certainty weighting. They further suggest, cautiously, that RLHF and instruction tuning may calibrate models toward conservative utility under uncertainty, while explicitly acknowledging this interpretation is speculative given the observational design. A key claim—that surface linguistic cues can override formal logical equivalence—is well supported by the data but rests on only two framings of one task.

The practical implication drawn is that stable cooperation among non-communicating agents is unlikely to emerge without explicit framing toward collective goals; embedding shared objectives in system prompts combined with ex-ante commitment mechanisms is proposed as a mitigation.

Limitations and open questions

The paper concedes several limitations directly. It is an observational study that cannot establish causal links between alignment methods and framing sensitivity; differences across families may reflect training data composition or architecture rather than alignment strategy. With $N=5$3 trials per model, statistical power is limited, particularly for smaller families, and chi-square expected-cell assumptions may be violated where Fisher's exact test would be more appropriate. The survival-based task is stylized—LLMs lack intrinsic survival drives—and results may not generalize to other framings or contexts. The independent-agent constraint also excludes coordination strategies available in interactive settings, so the claimed amplification relative to interactive multi-agent scenarios (H1's comparative component) is asserted rather than measured; no interactive baseline was run. Whether framing effects persist, attenuate, or reverse over multi-turn interactions remains an open question the paper explicitly defers to future work.

Conclusion

This study provides a cross-family behavioral analysis demonstrating that logically equivalent prompt framings produce large, statistically significant shifts in LLM decisions under complete agent isolation, with effect magnitudes ranging from 0.18 (Grok) to 0.90 (Claude). The findings establish framing as a potent bias source in non-interacting multi-agent deployments and motivate framing-aware prompt design and family-specific calibration. The causal origins of the observed risk-averse bias—whether alignment procedures, training distributions, or architectural factors—and the interaction of framing with actual inter-agent communication remain unresolved questions for subsequent controlled experimentation.