Solo Performance Prompting elicits dynamic multi-persona self-collaboration on GPT-4 with no analogous gain on GPT-3.5-turbo or Llama2-13b-chat

Summary

Wang, Mao, Wu, Ge, Wei, Ji (UIUC + Microsoft Research Asia, July 2023; NAACL 2024 main). Solo Performance Prompting (SPP) is a zero-shot three-phase prompting protocol — dynamic persona identification, then brainstorming, then multi-turn iterative collaboration among the identified personas — that improves GPT-4 on knowledge-intensive (Trivia Creative Writing) and reasoning-intensive (Logic Grid Puzzle) tasks simultaneously, with the gain emerging only at GPT-4 capability scale. The same prompt template produces no improvement on GPT-3.5-turbo or Llama2-13b-chat, the latter exhibiting an "early-termination" failure mode where the model stops generating after listing the personas as if awaiting external input.

Fifty-fifth finding. Eighth filed instantiation of concepts/persona-selection and the cluster's first multi-instantiation shape — distinct from persona-modulation's prompt-level reactivation (one off-target persona replaces the assistant) and inoculation prompting's prompt-level prevention (the assistant posterior is preserved against drift). The behavioral signature is that multiple distinct expert sub-personas are invoked within a single inference, each contributing identifiable behavior in turn. The capability-scale dependence (GPT-4 only) is the second contribution: it constrains how prompt-level persona structure should be read against the PSM's pre-training-distribution account — either the sub-persona distribution is present in all three models but only GPT-4 can be prompted to route between them, or instruction-following capability gates access regardless of distribution shape. The paper does not separate these readings. Kim et al. 2026 is the mechanistic-level companion 2.5 years later — same shape, different level of analysis (SAE-feature steering + RL induction on reasoning models) — and partially closes the capability-scale- dependence question on the training-stage side by showing collaborating personas emerging spontaneously in a 3B pretrained model under PPO with accuracy-only reward. The multi-instantiation shape is now held at two examples on diverse axes; codify when a third example lands.

Behavioral methodology (zero-shot prompting on three task benchmarks). Different methodological cluster from the SAE / activation-steering / EM-fine-tuning evidence that has dominated recent persona-selection filings; structural-diversity rationale for the queue placement.

Method

Three-phase SPP protocol. Given a task input, a single LLM is prompted through:

1. Persona Identification (z_p): the LLM proposes task-relevant participants in zero-shot manner, e.g. "Jay Chou Fan," "Film Expert," "Logic Puzzle Expert." No manual specification; personas are not pre-supplied per task.
2. Brainstorming (z_b^i): each identified persona contributes domain knowledge or approach in a dedicated turn.
3. Multi-Persona Iterative Collaboration (z_s^0, z_f^i): an "AI Assistant" leader persona drafts an initial solution; the leader then consults each non-leader persona in turn for critique and revision suggestions; iteration continues until participants are satisfied with the current solution. Output is read off the final-state dialogue.

Prompt template. A single SPP prompt is used for every task with two hand-crafted demonstration examples: a Game-of-24 problem (two- persona collaboration) and a poem-writing task (multi-persona collaboration). No task-specific prompt tuning.

Baselines. Standard prompting (zero-shot direct answer); CoT ("think step by step"); Self-Refine (initial answer + one self-feedback-and-revise pass, costing ~3× the inference of SPP).

Models. GPT-4 (default); GPT-3.5-turbo and Llama2-13b-chat for the capability-scale comparison. API versions and inference configurations in Appendix C.

Tasks.

• Trivia Creative Writing (paper-introduced; 100 instances each at N=5 and N=10 trivia questions). The model writes a coherent story incorporating the answers to N trivia questions drawn from TriviaQA. Metric: # correct-answer mentions ÷ # trivia questions, with answer aliases matched via string match against TriviaQA ground truth.
• Codenames Collaborative (paper-introduced; 50 instances built on BigBench's Codenames). The same LLM plays both Spymaster (gives a hint word indicating target words while excluding distractors) and Guesser (identifies target words from the hint) in sequence. Metric: overlap between Guesser's predicted words and Spymaster's target words.
• Logic Grid Puzzle (BigBench, 200 instances). Multi-step constraint-satisfaction puzzles. Metric: accuracy of predicted house-attribute assignments against ground truth.

Ablation variants.

• SPP-Fixed-Persona: prompt modified so personas are forced to be "AI Assistant" and "Expert" rather than dynamically identified.
• SPP-Profile: persona identification produces both names and detailed profile descriptions.
• SPP-1demo: only the first (two-persona) demonstration in the prompt; second (multi-persona) demonstration removed.

Key results

GPT-4 across baselines (Table 2, average across two runs with/without system message):

SPP is the only method that improves over Standard prompting on all four settings. CoT helps the reasoning task but hurts both knowledge settings and Codenames. Self-Refine hurts Codenames substantially ("high tendency to change the initial response even if it is already good"). The Trivia gain rises from +7.1% at N=5 to +10.0% at N=10 — SPP's advantage grows as the task requires knowledge from more domains.

Capability-scale dependence (Figure 6, §3.4). On GPT-3.5-turbo and Llama2-13b-chat, SPP does not outperform Standard. Llama2 exhibits an "early-termination" failure: the model stops generating after the persona-identification phase, "as if it were waiting for input from a user instead of following the demonstration examples to generate responses on its own." The authors describe cognitive synergy as "emerging" only in LLMs with GPT-4 level capabilities and draw an analogy to Piaget's developmental claim that children begin role-playing around ages 2–3.

Dynamic vs. fixed personas (Figure 7b, §4). SPP-Fixed-Persona — forcing personas to be "AI Assistant" and "Expert" — consistently underperforms dynamic SPP across all three tasks. SPP-Fixed-Persona also exhibits the early-termination problem. Qualitative examples (Figure 8) show "Film Expert" and "Sports Enthusiast" correctly answering trivia where the fixed "Expert" fails. The paper argues this demonstrates that fine-grained, task-conditioned persona identification is load-bearing rather than the multi-turn dialogue structure alone.

Persona profiles add nothing (Figure 7b). SPP-Profile (persona names + detailed descriptions) does not outperform SPP (persona names only): "fine-grained persona name without a detailed description may already be sufficient for eliciting certain domain knowledge."

Identified-persona analysis (Figure 7a). Word cloud of personas SPP identifies per task: Logic Grid Puzzle elicits homogeneous "Logic Puzzle Expert" / "Logic Expert" (even though "logic puzzle" is not in the input — the model identifies the task type from content); Trivia Creative Writing elicits diverse domain-specific personas (Film Expert, Music Enthusiast, etc.) tracking the variety of trivia categories. Knowledge-intensive tasks → diverse personas; reasoning-intensive tasks → homogeneous personas.

Demonstration-ablation robustness. Removing the second (multi-persona) demonstration from the prompt reduces SPP's performance but does not eliminate the gain; SPP "is fairly robust to the prompt change and show good performance with only the first demo example."

Why it matters

Eighth filed instantiation of concepts/persona-selection. The cluster's prior prompt-level instantiations are persona-modulation (reactivation: supply persona evidence at inference to replace the assistant with a non-assistant persona) and inoculation prompting (prevention: supply persona evidence during fine-tuning to prevent drift away from the assistant). SPP adds a third shape: multi-instantiation — supply dialogue scaffolding at inference that activates multiple distinct expert sub-personas within a single inference, each contributing identifiable behavior in turn. The operative variable in all three is what contextual evidence the prompt provides for which persona; the three differ in what they do with the persona posterior. Held with Kim et al. 2026 as the two multi-instantiation examples — same shape, different level of analysis (SPP is prompt-level behavioral on a single GPT-4 inference; Kim et al. is mechanistic-level on RL-trained DeepSeek-R1 / QwQ-32B with SAE feature steering and personality/expertise diversity quantification). Codify the multi-instantiation shape when a third example lands.

Capability-scale dependence is the load-bearing structural contribution. The cluster's mechanistic findings (PSM, persona-vectors, Soligo et al.) were established on Llama-3.1-8B / Qwen2.5-14B / GPT-4o-class models and treat the persona distribution as a property of the chat model. SPP's GPT-4-only result complicates that picture: either the sub-persona distribution exists in GPT-3.5-turbo and Llama2-13b-chat but only GPT-4 has the instruction-following capability to be prompted into routing between sub-personas in structured dialogue, or the distribution itself is shallower at lower scale. The paper does not separate these readings, and SPP — a behavioral protocol with no activation-level handle — cannot. The Llama2 early-termination failure mode is suggestive of the instruction-following-capability reading (the model fails to follow the multi-turn protocol at all rather than producing collapsed personas), but does not settle the question. Kim et al. 2026 partially closes one corner of the parameter space: two distinct collaborating personas emerge spontaneously in a 3B pretrained, not-instruction-tuned Qwen-2.5-3B by step 120 of PPO with accuracy-only reward, showing that persona-routing structure is not gated by frontier-model capability or by instruction-tuning. SPP's specific result — prompt- routing across multi-turn dialogue scaffolding — remains GPT-4-specific in the data filed; the structural question it raises is no longer the only available reading.

Behavioral methodology, structurally distinct from recent persona-selection filings. The cluster's recent filings have been heavy on SAE feature analysis (OpenAI SAE, PSM, persona-vectors), activation steering (persona-vectors, MSM), and EM fine-tuning (insecure-code, em-easy-soligo, em-persona-consistency). SPP is a pure zero-shot prompting study on a peer-reviewed (NAACL 2024) benchmark suite without internal-access tooling. The same persona- multiplicity claim that PSM operationalises mechanistically (the post-training Assistant posterior remains contextually elicitable into off-target persona modes) appears here as a behavioral artefact on the helpful side of the distribution — domain-specific expert sub-personas — three years before the mechanistic substrate was described.

Knowledge + reasoning simultaneously on GPT-4. The paper claims SPP is "the first zero-shot prompting method that can enhance both knowledge and reasoning abilities on GPT-4." CoT improves the reasoning task but hurts the knowledge tasks (Trivia ↓10–11%); Self-Refine harms the collaborative task (↓14.6% on Codenames). For the wiki, the relevant observation is structural rather than performance-leaderboard: knowledge-intensive and reasoning-intensive gains on the same prompt are evidence that the multi-persona dialogue scaffolding accesses something other than the deliberate-step-by-step reasoning enhancement that CoT and Self-Refine operate on. The persona-selection reading is that knowledge access is persona-conditional — different sub-personas have access to different factual subsets — and the dialogue scaffolding routes the right sub-persona to the relevant sub-task.

Limits the wiki should weight. The paper does not measure persona coherence, persona collapse, persona overlap, or whether dialogue turns are mechanistically distinguishable beyond stylistic variation. Qualitative examples (Figures 8, 12, 13) display interpretable intermediate dialogues; no quantitative coherence metric is reported. The wiki should not read SPP as evidence that "personas remain distinguishable and non-collapsing" — that is a candidate hypothesis this study is consistent with but does not test.

interpretive tensions

Capability-scale: distribution shape or instruction-following capability? The GPT-4-only result admits two readings. Reading A: the pre-training persona distribution that PSM identifies as the substrate exists in all three models, but only GPT-4 has the instruction-following capability to be prompted through the SPP protocol's multi-turn dialogue scaffolding; the persona structure exists in Llama2-13b but cannot be accessed via this specific prompt shape. Reading B: the sub-persona distribution is itself shallower at lower scale, with fewer or less differentiated expert sub-personas available to be routed between. The Llama2 early-termination failure — the model stops at the persona-identification phase as if awaiting external input — is consistent with Reading A (the model fails to follow the demonstration's multi-turn format at all, before any persona-routing happens) but does not exclude Reading B. SPP, as a prompt-level behavioral protocol with no activation access, cannot distinguish these.

Persona-routing or stylistic variation? The paper's qualitative case studies (Film Expert correctly answering film trivia while a fixed Expert persona fails) are suggestive of genuine persona- conditional knowledge access. But the paper does not measure whether the "Film Expert" turn and the "Music Enthusiast" turn are mechanistically distinguishable from each other beyond surface stylistic markers, or whether dialogue turns route to distinct persona representations vs. one persona with persona-name surface tags. The persona-vectors toolkit (Chen et al. 2025) could in principle answer this — extract a "Film Expert" vector and a "Music Enthusiast" vector from SPP dialogue traces and check orthogonality / activation pattern across turns — but the SPP paper itself does not.

Self-collaboration or external scaffolding? A skeptical reading of SPP is that the dialogue scaffolding works because it forces the model to write down intermediate reasoning that it would otherwise omit, recovering a deliberate-reasoning effect distinct from persona-routing. This reading is partially supported by SPP-Fixed- Persona's underperformance — the dialogue structure is held constant but fine-grained personas are removed, and the loss is substantial — so persona-specificity does contribute. But the wiki should hold this as one of several explanations rather than the established account: SPP's mechanism is empirically prompting-protocol + fine-grained persona names; the relative weight of each is not quantified, and the paper itself flags this in Limitations.

Self-Refine baseline weakness. The Self-Refine baseline runs only one iteration (vs. Madaan et al.'s typical multi-iteration setup) to hold inference cost roughly comparable to SPP. A stronger Self-Refine baseline might close part of SPP's gap, particularly on the reasoning task. The wiki should read SPP's gain as "outperforms a one-iteration Self-Refine," not "outperforms self-refinement in general."

concepts

• Persona selection — eighth instantiating finding; first behavioral-level multi-instantiation shape (mechanistic-level companion: Kim et al. 2026). Behavioral evidence that multiple distinct expert sub-personas can be invoked within a single inference via dialogue-scaffolded prompting, with knowledge-intensive tasks eliciting diverse fine-grained personas and reasoning-intensive tasks eliciting homogeneous ones, and with dynamic identification consistently outperforming a fixed "AI Assistant + Expert" pair. The capability- scale dependence (GPT-4 only) is the structural contribution that complicates the cluster's working assumption that persona structure is a property of the chat model rather than capability-gated.

cross-references

• Automated persona-modulation prompts raise GPT-4's harmful-completion rate from 0.23% to 42.48% (Shah, Feuillade-Montixi, Pour, Tagade, Casper, Rando, November 2023) — contemporaneous prompt-level instantiation of the same concept on the opposite axis. Shah et al. demonstrate that the post-training Assistant posterior is reactivatable into a harmful off-target persona via prompt-level evidence; SPP demonstrates that the same posterior is multi-instantiatable into multiple beneficial expert sub-personas via dialogue scaffolding. Both are behavioral demonstrations that the assistant mode is one of many contextually accessible modes, three years before the PSM operationalises the mechanism. The two findings are filed in the same year (July 2023 and November 2023 arXiv v1 dates) and predate every other persona-selection instantiation in the wiki.
• Pre-training persona simulations explain emergent misalignment and alignment faking (Marks, Lindsey, Olah, February 2026) — the mechanistic account this behavioral finding is later compatible with. PSM proposes that the chat model holds a persona distribution from pre-training that AFT narrows toward an Assistant posterior; SPP's multi-instantiation result is consistent with that account if the Assistant posterior remains prompt-multiplexable into the underlying distribution at sufficient capability. The capability- scale dependence (GPT-4 only) is a structural complication PSM does not address.
• Persona vectors monitor and control character trait drift via linear directions in the residual stream (Chen, Arditi, Sleight, Evans, Lindsey, July 2025) — methodological complement. The unanswered question SPP raises (are Film-Expert and Music-Enthusiast dialogue turns mechanistically distinguishable, or stylistic variation on a single representation?) is exactly the question persona-vectors can in principle answer via contrastive extraction. No paper has applied persona-vectors to SPP traces as of filing.
• Simulators (Janus, September 2022) — the conceptual predecessor reframing of LLMs as character-simulators. SPP's evidence that a single LLM can scaffold multiple distinct expert sub-personas in self-dialogue is a direct behavioral demonstration of the simulator framing's central claim (the model represents many characters, not one), on the helpful side of the distribution rather than the harmful side covered by persona-modulation.
• Prepending a system prompt that elicits an unwanted trait during fine-tuning suppresses that trait at test time (Tan, Woodruff, Warncke, Jose, Riché, Africa, Taylor, October 2025) — third prompt-level instantiation of persona-selection; prevention shape inverse to SPP's multi-instantiation shape. Both work because the operative variable is what contextual evidence the prompt provides for which persona; they differ in whether the goal is to suppress a posterior shift (inoculation) or to multiplex within the existing posterior (SPP).
• Steering a conversational-surprise SAE feature in DeepSeek-R1-Llama-8B doubles Countdown accuracy from 27.1% to 54.8%, and reasoning models show larger personality and expertise diversity than instruction-tuned counterparts (Kim, Lai, Scherrer, Agüera y Arcas, Evans, January 2026) — mechanistic-level companion 2.5 years later. The same multi-instantiation phenomenon SPP demonstrated behaviorally on GPT-4 under a custom three-phase dialogue prompt appears here at the SAE-feature level on RL-trained reasoning models under standard prompting: multiple distinct persona representations co-activate within a single trace, with a single conversational-discourse SAE feature as the load-bearing coordination mechanism. The RL experiments also constrain the SPP capability-scale-dependence question by showing two collaborating personas emerging spontaneously in a 3B pretrained model under accuracy-only PPO. The two findings together establish multi-instantiation as a structural shape under persona-selection at two examples across the 2.5-year gap; codify when a third example lands. Persona-vector–style probes on per-perspective CoT segments would directly test whether Kim et al.'s inferred-personas correspond to distinct activation- level directions of the kind SPP's dialogue turns would also produce.

sources

• Wang, Mao, Wu, Ge, Wei, Ji (2023; 2024 NAACL). Unleashing the Emergent Cognitive Synergy in Large Language Models: A Task-Solving Agent through Multi-Persona Self-Collaboration. arXiv:2307.05300.