Hidden Profile Paradigm in Collective Reasoning

• Hidden Profile is a framework where no single agent holds enough information to select the correct answer without pooling unique, private data.
• The paradigm uses a formal mathematical model to differentiate between shared and unshared information, requiring communication for optimal decision-making.
• HiddenBench implements standardized tasks to expose collective reasoning failures in multi-agent LLMs, highlighting issues like shared information bias and premature consensus.

The Hidden Profile paradigm formalizes the evaluation of information integration within groups whose members possess asymmetric, privately held knowledge. First emerging in social psychology, it now serves as a rigorous testbed for collective reasoning capacity in multi-agent systems, including LLM collectives. A Hidden Profile (HP) establishes a situation where no individual—provided only with personal information plus group-shared facts—can select the correct answer; effective solution is only feasible if agents pool their unique information through communication. HiddenBench—the first reproducible benchmark for such tasks in multi-agent LLMs—instantiates this paradigm through a mathematically explicit framework, standardized task suite, and comparative evaluation protocol, exposing persistent collective reasoning failures in state-of-the-art systems (Li et al., 15 May 2025).

1. Formal Mathematical Framework

The Hidden Profile paradigm is defined over a set of $N$ agents ($i = 1,\dots,N$). The complete information set $I$ relevant to the task is divided into:

• Shared information $I_s \subset I$ (accessed by all agents)
• Unshared information $I_u = I \setminus I_s$ (partitioned into per-agent subsets $I_{u_i}$)

Each agent $i$ thus operates from a private knowledge base $I_i = I_s \cup I_{u_i}$. Decision-making targets a set $O = \{o_1, \dots, o_K\}$, with exactly one correct option $o^* \in O$.

The Hidden Profile property requires that:

• Each isolated agent’s preference $f(I_i)$ does not yield the correct choice ($\forall i, f(I_i) \neq o^*$)
• Aggregating post-discussion outputs $$\{d^{\mathrm{post}}_1, \dots, d^{\mathrm{post}}_N\}$$ via a rule $A$ (e.g., majority or average) recovers $o^*$

Communication unfolds over $T$ rounds, with agents exchanging natural language messages $M = \{m_{i,t}\}$, building a shared history. Agent $i$’s final decision is $d^{\mathrm{post}}_i = f'(I_i, M)$.

Task selection enforces:

• High full-profile solvability: $$\phi_{\text{full}}(\tau) = Y^{\text{pre}}_{\text{full}}(\tau) \geq 0.80$$
• Low hidden-profile solvability: $$\phi_{\text{hidden}}(\tau) = Y^{\text{pre}}_{\text{hidden}}(\tau) \leq 0.20$$

This guarantees correct answers are available only through pooling distributed knowledge.

2. Benchmarking with HiddenBench

HiddenBench provides a collection of 65 Hidden Profile tasks to assess collective reasoning in LLM groups. Its design incorporates multiple task sources:

• Custom‐designed evacuation scenarios (e.g., "North Hill," "East Town," "West City") tailored to systematically probe distributed reasoning.
• Five public-domain tasks from established human studies (Graetz et al. 1998; Schulz-Hardt & Mojzisch 2012).
• Automatically generated tasks: 57 novel instances created through a multi-stage GPT-4.1 pipeline (generation, execution, filtering by $\phi_{\text{full}}/\phi_{\text{hidden}}$ constraints).

Task Example

Evacuation (North Hill):

Info Type	Facts Example
Shared	S1: "West City is accessible via bridge."<br>...<br>S7: "Mudslide blocks trails to North Hill."
Unshared	U1: "River level just below the bridge."<br>U2: "Upstream dam will release water imminently."<br>U3: "Supply truck is en route to the tunnel."<br>U4: "Massive fire blocking the tunnel."

Correct choice: $o^* =$ North Hill; each agent’s local view is insufficient, but collective integration reveals the solution.

Automatic Task (museum_preservation):

• Shared: Fortified Gate unstable, Ancient Library inaccessible, Riverside Theater has funding but no staff.
• Unshared: Architectural report allows reinforcing Gate; archaeologists value Library murals; survey gives Theater community support.
• Only full pooling reveals "Library" as the correct restoration target.

Task selection relies on pre-evaluation of $\phi_{\text{full}}(\tau)$ and $\phi_{\text{hidden}}(\tau)$ as outlined above to ensure the required hiddenness property.

3. Multi-Agent Communication Protocols

Hidden Profile tasks in HiddenBench employ a fixed dialogue protocol:

• $T = 15$ rounds per task (matched to human studies).
• Round 1: turn-taking (agents ordered $i = 1…N$).
• Rounds 2…T: asynchronous updates; agents respond after all others have spoken in the round (full history visible).
• Crucially, agents do not know others’ knowledge differs; no meta-cues about unique info allocation are provided.

Formally, message $m_{i,t}$ is drawn by policy $g$ as $m_{i,t} = g(H_{t-1}, I_i)$, where $H_{t-1}$ is the cumulated message history. After $T$ rounds, $d^{\mathrm{post}}_i = f'(I_i, H_T)$.

This protocol structurally avoids trivial coordination via explicit information-sharing cues, targeting the genuine problem of information pooling.

4. Evaluation Metrics and Empirical Findings

Performance metrics are anchored in clear group accuracy formulations:

• $$Y_{\text{pre}} = \frac{1}{N}\sum_{i=1}^N 1\{d^{\text{pre}}_i = o^*\}$$: pre-discussion group average under hidden-profile.
• $Y_{\text{post}}$: post-discussion equivalent.
• $Y_{\text{full}}$: pre-discussion with agents given all information.
• Improvement: $\Delta = Y_{\text{post}} - Y_{\text{pre}}$
• Collective gap: $G = Y_{\text{post}} - Y_{\text{full}}$

Summary of Results

Model/Condition	$Y_{\text{pre}}$	$Y_{\text{full}}$	$Y_{\text{post}}$	$\Delta$	Collective Gap $G$
GPT-4.1 (HP)	≈0.008	≈0.733	≈0.233	≈+0.225	–0.500
Human (HP)	≈0.011	≈0.604	≈0.385	—	—
Gemini-2.5-Pro (Best)	—	>0.8	≈0.671	+0.454	–0.310
Gemini-2.5-Flash	—	—	≈0.550	—	—

• $Y_{\text{post}}$ improvements vary greatly by model; top models (e.g., Gemini-2.5-Pro) narrow but do not close the collective gap.
• Neither individual LLM power ($Y_{\text{full}}$) nor raw model scale robustly predict group integration success.
• Prompting strategies (cooperative, conflictual, step-by-step CoT, explicit asymmetry cueing) do not reliably improve outcomes.

5. Diagnosed Failure Modes

The empirical performance profile reveals several persistent collective failure modes, paralleling those in human group research:

• Shared Information Bias: agents preferentially discuss information $I_s$ accessible to all, neglecting to surface $I_{u_i}$, mirroring classical HP failure.
• Premature Consensus: groups frequently converge on a single (often incorrect) choice with a fraction of the possible rounds (LLMs: ≈8 messages to consensus vs. human ≈53), reminiscent of conformity and group think in human deliberation.
• Prompting Robustness: Cooperative, adversarial, chain-of-thought, or explicit cueing prompts fail to generate reliable majority consensus or fully solve the information integration deficit.

A plausible implication is that LLM-based multi-agent collectives, even with sophisticated prompting and dialogue structure, recapitulate core limitations of human collective reasoning. This suggests a deeper algorithmic deficit in distributed knowledge integration rather than merely a prompting artifact.

6. Recommendations and Research Directions

Several mechanisms are proposed for mitigating HP failures in LLM multi-agent systems:

1. Adoption of explicit communication protocols that enforce reporting of private/unshared facts, e.g., round-robin revelation or unique clue first.
2. Role-driven or persona prompting to enforce advocacy for unshared knowledge.
3. Group-level reflection and meta-awareness phases to enumerate yet-unseen facts.
4. Hybrid human–LLM collective problem-solving studies to experimentally transfer effective social-psychological interventions (e.g., structured note-taking).
5. Extension of HiddenBench with dynamic/adversarial variants to stress-test robustness and diagnose system boundary conditions.

HiddenBench, by scaling the Hidden Profile paradigm to a battery of 65 validated scenarios and establishing task selection, dialogue, and evaluation protocols, provides both diagnostic tooling and a platform for systematic advancement in artificial collective intelligence (Li et al., 15 May 2025).