betaFeedback

WorksComputer Science

Stream of Belief: Eliciting Collective Intelligence through Latent Belief Dynamics

Abstract

In this work, we view multi-agent reasoning as a process of seeking a collective equilibrium of implicit belief dynamics and introduce Stream of Belief (SoB), where agents maintain continuous latent beliefs and iteratively refine them through communication and reflection. Our method seeks a collective belief state of a multi-agent system where further communication induces no meaningful information propagation. Through extensive experiments, we demonstrate the generalization of SoB across tasks and model families. More Broadly, our results suggest a new perspective on scaling multi-agent reasoning from belief dynamics.

Embodied-Minds-Lab/Stream-of-Belief

Introduction

Collaboration among language agents is becoming an important paradigm for collective intelligence in the era of Agentic AI [5, 36, 72]. Recent advances in multi-agent systems (MAS) [41, 43, 48, 79] have extended reasoning from individual models to collaborative systems of interacting agents, enabling applications in mathematical and scientific reasoning [24, 33, 51, 68, 95], automated code generation [45, 58, 71], and open-domain question answering [12, 50, 91].

This paradigm echoes a similar pattern in human societies: collective intelligence. In human societies, individuals with incomplete knowledge interact with each other and continually revise their beliefs through interaction, enabling collective capabilities that exceed those of isolated individuals [19, 29, 40, 66, 78]. Classical studies of social influence and opinion dynamics model how individuals revise their beliefs in response to others, revealing how local interactions shape the evolution of collective states [10, 20, 21, 27, 38]. Subsequent work on social learning further shows that the structure of the interactions determines the final conclusions [7, 32, 46, 54]. These studies motivate us to rethink collective reasoning through how individual beliefs are represented and updated through interaction.

However, translating this belief dynamics perspective into language-model agents raises two fundamental questions: how should an agent’s belief be represented, and how should interactions update these beliefs? Existing multi-agent systems primarily rely on natural-language communication, where agents exchange, critique, and revise explicit reasoning traces through dialogue [12, 24, 52]. Recent advances in latent reasoning have explored continuous representations for intermediate computation, either by recurrently feeding back hidden states [30, 37, 62] or by constructing mixture token embeddings [63, 80, 92]. Extending this direction to multi-agent systems, latent communication methods enable agents to exchange internal representations through activations, hidden states, or key–value caches [28, 60, 86], with recent frameworks supporting latent collaboration through shared working memories [98], learned communication prefixes [94], or trained receivers [25]. Together, these advances establish continuous representations as a medium for reasoning and communication, motivating a further question:

Can collective intelligence emerge from interacting latent beliefs?

Figure 1. Illustration of multi-agent communication from a belief dynamics view. Schematic: the landscapes are illustrative; the axes are two coordinates of the latent belief space, and the basins are basins of attraction of the belief update.

Motivated by this perspective, we formulate multi-agent reasoning as the evolution of a joint belief state and introduce Stream of Belief (SoB). We propose to construct Belief Tokens, an agentic communication medium, to facilitate the belief propagation of multi-agent systems in high-dimensional latent space. Additionally, in order to push forward the joint belief state, we construct two complementary processes: local refinement, through which each agent independently refines its beliefs, and global interaction, through which information propagates across agents. The interplay between local and global operations allows individual belief trajectories to continually reshape one another.

Stream of Belief iterates the joint belief state toward a stable equilibrium, at which further interaction induces no meaningful information propagation, and decodes a collective solution from the resulting latent state. Equilibrium does not imply consensus: agents may retain heterogeneous beliefs while achieving global stability, as in classical influence models whose stable states retain persistent disagreement [1, 4, 8, 27].

Our contributions are threefold. (1), we formulate collective reasoning as the evolution of interacting latent beliefs toward equilibrium. (2), we propose Stream of Belief, a multi-agent collaboration framework that outperforms the strongest latent reasoning baselines, (i) improving accuracy by up to 15.7 points, (ii) reducing text-token usage by 51% on average (up to 82%), and (iii) delivering 1.1×–2.6× lower token latency than the most efficient latent baselines. (3), through extensive experiments, we demonstrate the generalization of our method across tasks and models.

Preliminaries and Notations

Language model. Let fθf_{\theta} denote a frozen autoregressive language model with vocabulary V\mathcal{V} and input embedding matrix EE, where E[v]∈RdE[v] \in\mathbb{R}^{d} is the embedding of token vv. Given an input sequence, the model produces a last-layer hidden state hh and predicts the next-token distribution p=softmax⁡(Wouth)p = \operatorname{softmax}(W_{\mathrm{out}}h). The autoregressive generation selects a discrete token from pp and feeds its embedding back as the next input.

Belief tokens. Following the probability-weighted embedding formulation of Soft Thinking [92], we represent each generation step by a continuous mixture of token embeddings. Given the next-token distribution pℓp_{\ell} at generation step ℓ\ell, the belief token is

zℓ=∑v∈Vpℓ(v)E[v].(1)z_{\ell} = \sum_{v \in\mathcal{V}} p_{\ell}(v)E[v]. \tag*{(1)}

Each belief token is fed back as the next input embedding.1 Repeating this process for LL autoregressive steps yields a belief block z=(z1,…,zL)∈RL×dz = (z_{1}, \ldots, z_{L}) \in\mathbb{R}^{L \times d}. By construction, belief tokens lie in the convex hull of vocabulary embeddings and can be fed directly into the model without a learned projection. We examine their distributional properties in § 5.1.

Example: Expressiveness of Belief Tokens.

Consider a multiple-choice question where the next-token distribution assigns probabilities to four answer labels AA, BB, CC, and DD. Suppose the normalized weights are

(p(A),p(B),p(C),p(D))=(0.55,0.30,0.15,0.00).\left(p(A), p(B), p(C), p(D)\right) = (0.55, 0.30, 0.15, 0.00).

Conventional greedy decoding selects AA and feeds back its embedding E[A]E[A]. In contrast, a belief token represents a probability-weighted combination of multiple candidates:

z=0.55 E[A]+0.30 E[B]+0.15 E[C]+0.00 E[D].z = 0.55\,E[A] + 0.30\,E[B] + 0.15\,E[C] + 0.00\,E[D].

Our continuous representation incorporates contributions from multiple candidate embeddings. As agents exchange and refine their beliefs, subsequent token distributions may change, allowing different candidates to influence the evolving latent representations.

Agents and beliefs. We consider NN agents A1,…,ANA_{1}, \ldots, A_{N} instantiated from the same frozen model fθf_{\theta} with agent-specific personas.2 We write Ai(c)∈RL×dA_{i}(c) \in\mathbb{R}^{L \times d} for the block of LL belief tokens generated autoregressively by agent ii from context cc using Eq. (1). The context combines text inputs (the problem and instructions) with continuous belief blocks inserted directly as input embeddings (Figure 2). We refer to each agent’s current latent reasoning state zi∈RL×dz_{i} \in\mathbb{R}^{L \times d} as its belief. The collection of agent beliefs forms the joint state Z=(z1,…,zN)Z = (z_{1}, \ldots, z_{N}). Reasoning proceeds in cycles indexed by kk, with zi(k)z_{i}^{(k)} and Z(k)Z^{(k)} denoting the states at the start of cycle kk; the cycle index is omitted when describing an individual update.

Figure 2. Stream of Belief at the embedding layer. (a) Global interaction conditions each agent on its own and other agents’ beliefs, while (b) local refinement uses only its own belief. Both operations autoregressively generate LL continuous belief tokens via probability-weighted token embeddings (Eq. (1)), replacing the previous belief block.

Stream of Belief

Each agent initializes its belief from the problem xx as zi←Ai(x)z_i \leftarrow A_i(x). The beliefs are subsequently updated through two complementary operations:

Global Interaction. Each agent updates its belief by incorporating the current beliefs of other agents alongside its own, enabling information to propagate across the system:

[G(Z)]i=Ai(x,zi,{zj}j≠i).(2)[G(Z)]_i = A_i\left(x, z_i, \{z_j\}_{j \ne i}\right). \tag*{(2)}

Local Refinement. Each agent independently refines its belief by conditioning only on its previous belief and the problem:

[L(Z)]i=Ai(x,zi).(3)[L(Z)]_i = A_i\left(x, z_i\right). \tag*{(3)}

Both operations employ the same belief-generation procedure, differing only in the information available to each agent. All agents update synchronously, with newly generated belief blocks replacing their previous states while maintaining a fixed size of L×dL \times d (Figure 2). Each update reconstructs its context from the problem and current beliefs, preventing the accumulation of intermediate states across cycles.

Belief Update. We organize collective reasoning into cycles, each consisting of one global interaction followed by mm local refinements:

Z(k+1)=Φm(Z(k)),Φm=Lm∘G.(4)Z^{(k+1)} = \Phi_m(Z^{(k)}), \qquad\Phi_m = L^m \circ G. \tag*{(4)}

Here, mm controls the depth of independent refinement between successive interactions, with m=0m = 0 corresponding to interaction-only updates. We refer to the repeated application of Φm\Phi_m as belief iteration.

Algorithm 1 | Stream of Belief

Require: problem xx; agents {Ai}i=1N\{A_i\}_{i=1}^{N}; depth mm; patience PP; maximum cycles Kmax⁡K_{\max}

  1. Z←(A1(x),…,AN(x))Z \leftarrow(A_1(x), \ldots, A_N(x))

  2. for k=1,…,Kmax⁡k = 1, \ldots, K_{\max} do

  3. Z←Φm(Z)Z \leftarrow\Phi_m(Z) ▹ belief iteration

  4. y^(k)←Maj⁡iDecode⁡(x,zi)\hat{y}^{(k)} \leftarrow\operatorname{Maj}_i \operatorname{Decode}(x,z_i) ▹ short probes

  5. if k≥Pk \ge P and y^(k−P+1)=⋯=y^(k)\hat{y}^{(k-P+1)} = \cdots= \hat{y}^{(k)} then

  6. break ▹ early stopping

  7. return Maj⁡iDecode⁡(x,zi)\operatorname{Maj}_i \operatorname{Decode}(x,z_i) ▹ full solutions

Collective Equilibrium and Inference. Through repeated belief iteration, the joint state evolves under Φm\Phi_m. A collective equilibrium is a fixed point Z⋆=Φm(Z⋆)Z^\star= \Phi_m(Z^\star), where further interaction leaves the joint state unchanged. We characterize convergence through the relative belief residual between successive cycles:

rk=∥Z(k+1)−Z(k)∥F∥Z(k)∥F+δ,(5)r_k = \frac{\lVert Z^{(k+1)} - Z^{(k)} \rVert_F}{\lVert Z^{(k)} \rVert_F + \delta}, \tag*{(5)}

where δ>0\delta> 0 ensures numerical stability.

Intermediate Decoding and Early Stopping. We denote by Decode⁡(x,zi)\operatorname{Decode}(x,z_i) the textual answer decoded by agent ii conditioned on problem xx and its belief ziz_i, and by Maj⁡i\operatorname{Maj}_i the majority vote over agents’ decoded answers. After each cycle kk, agents decode short probe answers, whose majority defines the collective probe y^(k)\hat{y}^{(k)}. These probes are used only for monitoring and do not re-enter belief updates. Inference stops when the collective probe remains unchanged for PP consecutive cycles (patience) or reaches Kmax⁡K_{\max} cycles (maximum budget). Finally, each agent decodes a full solution from its final belief, and the collective prediction is obtained by majority voting. Probe and final decoding use the same model with different prompts and token budgets (Appendix § 10); Algorithm 1 summarizes the full procedure.

Experiments

Our experiments investigate four questions: 1. How well does SoB generalize across reasoning tasks, model families, and model scales (§ 4.3)? 2. How does SoB trade off reasoning accuracy against inference latency and token usage (§ 5)? 3. How do latent belief representations evolve relative to the model’s native embedding space (§ 5.1)? 4. How do belief capacity, agent population, and iteration dynamics affect collective reasoning (§ 5.2)?

Experimental Setup

Base models. We use one instruction-tuned model from each of four families: Qwen3-8B (non-thinking mode) [81], Llama-3.1-8B-Instruct [35], Gemma-3-4B-it [31], and Phi-4-mini-instruct [2]. For the scaling study (Table 2) we additionally use the Qwen3 dense series from 0.6B to 32B and the Llama family, Llama-3.2-1B/3B-Instruct, Llama-3.1-8B-Instruct, Llama-3.3-70B-Instruct and Llama-3.1-405B-Instruct.

ModelMethodMATH-500GPQA-DMedQABBH
Multi-agent debate57.2±3.257.2\pm3.223.2±2.923.2\pm2.934.0±2.934.0\pm2.935.2±2.335.2\pm2.3
Qwen3-0.6BLatentMAS14.4±2.114.4\pm2.114.1±2.414.1\pm2.416.3±2.116.3\pm2.143.5±2.543.5\pm2.5
SwiReasoning39.6±3.139.6\pm3.110.6±2.210.6\pm2.214.7±2.114.7\pm2.126.7±2.226.7\pm2.2
Soft Thinking26.0±2.826.0\pm2.83.5±1.33.5\pm1.312.0±2.012.0\pm2.024.0±2.124.0\pm2.1
Ours49.6±3.149.6\pm3.125.3±3.225.3\pm3.234.7±2.934.7\pm2.943.0±2.543.0\pm2.5
Multi-agent debate77.6±2.677.6\pm2.634.8±3.434.8\pm3.447.7±2.847.7\pm2.872.5±2.372.5\pm2.3
Qwen3-1.7BLatentMAS24.4±2.824.4\pm2.812.1±2.412.1\pm2.422.3±2.422.3\pm2.452.0±2.552.0\pm2.5
SwiReasoning44.0±3.144.0\pm3.115.7±2.615.7\pm2.636.7±2.736.7\pm2.759.2±2.559.2\pm2.5
Soft Thinking58.8±3.158.8\pm3.15.6±1.65.6\pm1.69.3±1.79.3\pm1.737.5±2.337.5\pm2.3
Ours66.4±3.166.4\pm3.133.3±3.433.3\pm3.447.3±2.947.3\pm2.975.2±2.375.2\pm2.3
Multi-agent debate86.4±2.286.4\pm2.245.5±3.545.5\pm3.559.7±2.959.7\pm2.992.0±1.392.0\pm1.3
Qwen3-4BLatentMAS50.8±3.150.8\pm3.126.8±3.126.8\pm3.149.0±3.049.0\pm3.083.0±1.983.0\pm1.9
SwiReasoning47.6±3.147.6\pm3.122.7±3.022.7\pm3.032.0±2.732.0\pm2.758.0±2.458.0\pm2.4
Soft Thinking62.4±3.262.4\pm3.29.1±2.09.1\pm2.06.7±1.46.7\pm1.415.0±1.815.0\pm1.8
Ours77.2±2.677.2\pm2.643.9±3.743.9\pm3.766.7±2.866.7\pm2.893.5±1.293.5\pm1.2
Multi-agent debate86.4±2.186.4\pm2.153.0±3.453.0\pm3.473.7±2.673.7\pm2.693.5±1.393.5\pm1.3
Qwen3-8BLatentMAS65.6±2.965.6\pm2.934.3±3.334.3\pm3.364.0±2.864.0\pm2.886.5±1.786.5\pm1.7
SwiReasoning47.6±3.247.6\pm3.226.3±3.126.3\pm3.144.7±2.844.7\pm2.854.0±2.554.0\pm2.5
Soft Thinking57.2±3.057.2\pm3.016.7±2.616.7\pm2.612.0±1.912.0\pm1.915.3±1.815.3\pm1.8
Ours77.2±2.777.2\pm2.744.9±3.544.9\pm3.576.3±2.476.3\pm2.493.5±1.293.5\pm1.2
Multi-agent debate85.6±2.385.6\pm2.352.5±3.552.5\pm3.574.7±2.574.7\pm2.597.0±0.997.0\pm0.9
Qwen3-14BLatentMAS75.2±2.775.2\pm2.742.4±3.542.4\pm3.579.7±2.379.7\pm2.392.2±1.392.2\pm1.3
SwiReasoning46.8±3.146.8\pm3.125.8±2.925.8\pm2.961.3±2.861.3\pm2.865.2±2.565.2\pm2.5
Soft Thinking54.8±3.254.8\pm3.218.2±2.718.2\pm2.714.3±1.914.3\pm1.923.5±2.223.5\pm2.2
Ours80.0±2.680.0\pm2.653.5±3.553.5\pm3.580.0±2.280.0\pm2.296.8±0.996.8\pm0.9
Multi-agent debate88.4±2.088.4\pm2.059.6±3.559.6\pm3.583.3±2.283.3\pm2.297.0±0.897.0\pm0.8
Qwen3-32BLatentMAS76.0±2.776.0\pm2.746.5±3.646.5\pm3.686.7±1.986.7\pm1.996.0±1.096.0\pm1.0
SwiReasoning48.4±3.248.4\pm3.230.3±3.330.3\pm3.375.3±2.575.3\pm2.578.0±2.178.0\pm2.1
Soft Thinking62.8±3.162.8\pm3.137.9±3.537.9\pm3.519.0±2.219.0\pm2.251.5±2.551.5\pm2.5
Ours82.0±2.482.0\pm2.451.5±3.551.5\pm3.584.7±2.184.7\pm2.197.5±0.897.5\pm0.8
ModelMethodMATH-500GPQA-DMedQABBH
Multi-agent debate28.0±2.828.0\pm2.817.2±2.717.2\pm2.726.3±2.526.3\pm2.526.2±2.226.2\pm2.2
Llama-3.2-1BLatentMAS17.6±2.417.6\pm2.415.2±2.615.2\pm2.623.7±2.423.7\pm2.410.2±1.510.2\pm1.5
SwiReasoning23.2±2.623.2\pm2.616.7±2.716.7\pm2.726.0±2.526.0\pm2.522.8±2.122.8\pm2.1
Soft Thinking16.4±2.316.4\pm2.38.6±2.08.6\pm2.024.7±2.424.7\pm2.420.8±2.020.8\pm2.0
Ours19.2±2.519.2\pm2.519.7±2.819.7\pm2.842.7±2.942.7\pm2.929.5±2.329.5\pm2.3
Multi-agent debate45.6±3.145.6\pm3.130.3±3.230.3\pm3.260.0±2.860.0\pm2.849.2±2.549.2\pm2.5
Llama-3.2-3BLatentMAS35.2±3.035.2\pm3.020.7±2.820.7\pm2.852.7±2.952.7\pm2.945.8±2.545.8\pm2.5
SwiReasoning38.8±3.138.8\pm3.125.3±3.125.3\pm3.148.7±2.948.7\pm2.945.0±2.445.0\pm2.4
Soft Thinking29.6±2.929.6\pm2.917.5±2.817.5\pm2.844.0±2.944.0\pm2.949.3±2.549.3\pm2.5
Ours34.0±3.134.0\pm3.131.3±3.431.3\pm3.456.0±2.956.0\pm2.948.0±2.548.0\pm2.5
Multi-agent debate54.0±3.054.0\pm3.030.3±3.130.3\pm3.170.0±2.670.0\pm2.660.5±2.460.5\pm2.4
Llama-3.1-8BLatentMAS46.0±3.446.0\pm3.422.2±2.922.2\pm2.968.3±2.668.3\pm2.666.2±2.366.2\pm2.3
SwiReasoning48.8±3.148.8\pm3.119.2±2.819.2\pm2.873.3±2.573.3\pm2.562.2±2.562.2\pm2.5
Soft Thinking39.6±3.139.6\pm3.123.7±3.123.7\pm3.163.7±2.863.7\pm2.865.5±2.265.5\pm2.2
Ours46.8±3.146.8\pm3.131.8±3.331.8\pm3.369.3±2.769.3\pm2.770.0±2.270.0\pm2.2
Multi-agent debate77.6±2.677.6\pm2.653.0±3.653.0\pm3.689.0±1.989.0\pm1.997.5±0.897.5\pm0.8
Llama-3.3-70BLatentMAS73.6±2.873.6\pm2.844.9±3.544.9\pm3.582.7±1.882.7\pm1.896.8±0.996.8\pm0.9
SwiReasoning72.8±2.972.8\pm2.948.0±3.448.0\pm3.488.3±1.988.3\pm1.995.8±1.095.8\pm1.0
Soft Thinking72.8±2.772.8\pm2.752.0±3.552.0\pm3.587.3±1.987.3\pm1.995.8±1.095.8\pm1.0
Ours78.8±2.678.8\pm2.656.6±3.556.6\pm3.585.7±2.085.7\pm2.096.8±0.996.8\pm0.9
Multi-agent debate75.6±2.775.6\pm2.751.5±3.651.5\pm3.686.7±1.986.7\pm1.997.5±0.897.5\pm0.8
Llama-3.1-405BLatentMAS68.4±3.068.4\pm3.049.0±3.549.0\pm3.584.0±2.184.0\pm2.197.2±0.897.2\pm0.8
SwiReasoning72.0±2.972.0\pm2.949.0±3.549.0\pm3.585.0±2.185.0\pm2.195.5±1.095.5\pm1.0
Soft Thinking71.2±3.071.2\pm3.046.0±3.646.0\pm3.687.0±1.987.0\pm1.997.2±0.897.2\pm0.8
Ours73.2±2.873.2\pm2.855.5±3.455.5\pm3.489.0±1.789.0\pm1.797.5±0.897.5\pm0.8

Table 2. Accuracy (%) across model scales.

Benchmarks. Eight benchmarks cover mathematical, scientific, commonsense, logical, and multi-hop reasoning: GSM8K [17], MATH-500 [39, 53], MMLU-Pro [75], GPQA-Diamond [61], ARC-Challenge [16], MedQA [47], BIG-Bench Hard [67], and HotpotQA [83].

Accuracy

Figure 3. Qualitative demonstration of reaching a collective equilibrium. (a) Schematic: the beliefs of four agents on a two-dimensional belief landscape, drawn so that each belief moves in the cycles in which its decoded answer in (b) changes (hollow markers: initial beliefs). In cycles 0–2 the agents take large steps towards an attractor, and in cycles 3–6 they make small adjustments inside its basin. Agents A, B, and D settle in the basin decoded as “yes” and agent C in the neighboring basin decoded as “no”. (b) A graph-connectivity problem with distributed evidence (Qwen3-8B, N=4N = 4; each agent observes only its own edges). The decoded answers are shown at every cycle together with the belief residual rkr_k. The decoded answers stop changing after cycle 3, while the latent beliefs keep moving until cycle 6, where r6<ϵr_6 < \epsilon: the joint state is stationary although the agents do not agree (3:1), and the collective answer is decoded from this equilibrium.

Protocol. N=4N=4 agents share the frozen base model and update synchronously. Each belief ziz_i is a block of L=384L=384 belief tokens. Each cycle is one global interaction followed by m=1m=1 local refinement; we run at most Kmax⁡=16K_{\max}=16 cycles and stop once the collective probe is unchanged for P=2P=2 cycles. Each agent then decodes a solution of up to 512 tokens from its own belief, and y^\hat{y} is the majority over parsable answers.

Baselines. We compare with six latent reasoning methods. LatentMAS [98], which passes latent thoughts through a shared working memory without training; ThoughtComm [94], which extracts latent thoughts from the agents’ hidden states with a sparse autoencoder and injects them into the other agents as trained prefixes; Interlat [25], which trains a receiver to read the last hidden states of a sender; SwiReasoning [63] and Soft Thinking [92], both training-free, and SoftCoT [80], which trains a projection from the soft thoughts of a small assistant model into the embedding space of the frozen base model. We also compare with multi-agent debate [24]. Implementation details of all baselines are given in Appendix § 11.

Accuracy

Observation 1. Interacting latent beliefs provide an effective and generalizable mechanism for collective reasoning across tasks and model families.

Table 1 shows that collective reasoning can be effectively organized through continuous latent interaction, consistently outperforming latent baselines across diverse tasks and models. Its competitive performance with text-based debate suggests effective collaboration can also emerge through latent interaction.

ModelMethodMath and scienceKnowledge, logic, and multi-hop
GSM8KMATH-500MMLU-ProGPQA-DARC-CMedQABBHHotpotQA
Qwen3-8BMulti-agent debate96.5±1.396.5\pm1.386.4±2.286.4\pm2.275.3±2.575.3\pm2.553.0±3.453.0\pm3.494.5±1.694.5\pm1.673.7±2.673.7\pm2.693.5±1.293.5\pm1.270.7±3.670.7\pm3.6
LatentMAS87.5±2.387.5\pm2.365.6±3.165.6\pm3.162.7±2.762.7\pm2.734.3±3.434.3\pm3.491.0±2.191.0\pm2.164.0±2.864.0\pm2.886.5±1.686.5\pm1.660.0±3.960.0\pm3.9
ThoughtComm43.5±3.543.5\pm3.55.2±1.45.2\pm1.416.0±2.216.0\pm2.213.1±2.513.1\pm2.569.5±3.369.5\pm3.333.3±2.733.3\pm2.724.8±2.024.8\pm2.048.7±4.248.7\pm4.2
Interlat84.5±2.584.5\pm2.547.2±3.247.2\pm3.225.7±2.425.7\pm2.413.1±2.413.1\pm2.485.0±2.585.0\pm2.552.7±3.052.7\pm3.056.8±2.456.8\pm2.459.3±4.059.3\pm4.0
SwiReasoning23.0±2.923.0\pm2.925.6±2.625.6\pm2.621.0±2.321.0\pm2.37.6±1.97.6\pm1.919.5±2.719.5\pm2.718.3±2.218.3\pm2.225.8±2.225.8\pm2.216.7±3.116.7\pm3.1
Soft Thinking5.5±1.65.5\pm1.623.2±2.723.2\pm2.77.7±1.57.7\pm1.51.5±0.91.5\pm0.911.0±2.211.0\pm2.24.3±1.24.3\pm1.25.5±1.15.5\pm1.130.0±3.730.0\pm3.7
SoftCoT39.0±3.539.0\pm3.526.0±2.726.0\pm2.713.0±2.013.0\pm2.09.6±2.19.6\pm2.114.5±2.514.5\pm2.58.3±1.68.3\pm1.624.8±1.824.8\pm1.817.3±3.117.3\pm3.1
Ours94.5±1.694.5\pm1.677.2±2.777.2\pm2.769.7±2.769.7\pm2.744.9±3.644.9\pm3.692.5±1.892.5\pm1.876.3±2.476.3\pm2.493.5±1.293.5\pm1.261.3±3.861.3\pm3.8
Phi-4-miniMulti-agent debate88.5±2.288.5\pm2.258.4±3.258.4\pm3.255.7±2.955.7\pm2.932.8±3.232.8\pm3.287.0±2.587.0\pm2.555.0±3.055.0\pm3.081.2±2.081.2\pm2.024.0±3.524.0\pm3.5
LatentMAS73.5±3.073.5\pm3.011.2±1.911.2\pm1.938.3±2.738.3\pm2.723.7±3.023.7\pm3.079.0±3.079.0\pm3.025.3±2.525.3\pm2.562.0±2.462.0\pm2.446.0±4.046.0\pm4.0
ThoughtComm61.0±3.561.0\pm3.537.2±3.037.2\pm3.025.3±2.525.3\pm2.516.7±2.716.7\pm2.736.5±3.236.5\pm3.234.0±2.834.0\pm2.842.8±2.442.8\pm2.421.3±3.321.3\pm3.3
Interlat68.5±3.368.5\pm3.332.0±2.932.0\pm2.924.7±2.524.7\pm2.521.2±2.821.2\pm2.877.5±2.977.5\pm2.938.7±2.838.7\pm2.856.2±2.456.2\pm2.450.0±4.250.0\pm4.2
SwiReasoning83.5±2.683.5\pm2.651.6±3.251.6\pm3.237.7±2.737.7\pm2.711.6±2.311.6\pm2.370.0±3.170.0\pm3.154.0±2.954.0\pm2.963.5±2.363.5\pm2.320.0±3.320.0\pm3.3
Soft Thinking77.0±2.977.0\pm2.936.0±3.036.0\pm3.030.7±2.830.7\pm2.87.6±1.97.6\pm1.971.0±3.171.0\pm3.138.3±3.038.3\pm3.071.5±2.071.5\pm2.08.7±2.38.7\pm2.3
SoftCoT50.0±3.650.0\pm3.617.6±2.417.6\pm2.410.3±1.810.3\pm1.816.2±2.616.2\pm2.650.0±3.650.0\pm3.627.0±2.627.0\pm2.632.8±2.332.8\pm2.318.0±2.918.0\pm2.9
Ours91.0±2.191.0\pm2.156.0±3.356.0\pm3.354.0±2.954.0\pm2.928.3±3.228.3\pm3.290.5±2.290.5\pm2.255.3±2.855.3\pm2.880.2±1.880.2\pm1.837.3±3.937.3\pm3.9
Llama-3.1-8BMulti-agent debate89.5±2.189.5\pm2.154.0±3.254.0\pm3.247.7±2.847.7\pm2.830.3±3.330.3\pm3.383.5±2.683.5\pm2.670.0±2.670.0\pm2.660.5±2.560.5\pm2.551.3±4.051.3\pm4.0
LatentMAS85.5±2.585.5\pm2.546.0±3.046.0\pm3.040.0±3.040.0\pm3.022.2±3.022.2\pm3.079.5±2.979.5\pm2.968.3±2.768.3\pm2.766.2±2.366.2\pm2.367.3±3.767.3\pm3.7
ThoughtComm72.5±3.372.5\pm3.329.6±2.929.6\pm2.924.3±2.524.3\pm2.512.6±2.312.6\pm2.373.5±2.973.5\pm2.952.7±3.052.7\pm3.052.5±2.452.5\pm2.444.7±4.144.7\pm4.1
Interlat53.5±3.653.5\pm3.625.6±2.725.6\pm2.721.3±2.321.3\pm2.323.2±3.023.2\pm3.035.0±3.535.0\pm3.534.0±2.734.0\pm2.735.8±2.435.8\pm2.49.3±2.49.3\pm2.4
SwiReasoning87.0±2.387.0\pm2.349.6±3.149.6\pm3.142.7±2.842.7\pm2.824.2±3.124.2\pm3.182.5±2.782.5\pm2.769.0±2.769.0\pm2.760.0±2.460.0\pm2.455.3±4.255.3\pm4.2
Soft Thinking81.5±2.881.5\pm2.840.8±3.240.8\pm3.231.0±2.731.0\pm2.713.6±2.413.6\pm2.468.5±3.268.5\pm3.236.7±2.736.7\pm2.760.8±2.360.8\pm2.326.0±3.526.0\pm3.5
SoftCoT47.0±3.447.0\pm3.413.6±2.113.6\pm2.120.7±2.320.7\pm2.323.7±3.023.7\pm3.066.0±3.366.0\pm3.347.3±2.847.3\pm2.833.2±2.333.2\pm2.337.3±4.037.3\pm4.0
Ours84.0±2.684.0\pm2.646.8±3.146.8\pm3.147.7±2.947.7\pm2.931.8±3.331.8\pm3.384.5±2.684.5\pm2.669.3±2.769.3\pm2.770.0±2.370.0\pm2.351.3±4.151.3\pm4.1
Gemma-3-4BMulti-agent debate89.5±2.189.5\pm2.180.8±2.580.8\pm2.547.7±2.847.7\pm2.832.3±3.332.3\pm3.378.0±2.978.0\pm2.945.0±2.945.0\pm2.989.8±1.589.8\pm1.554.0±4.154.0\pm4.1
LatentMAS83.0±2.783.0\pm2.758.4±3.258.4\pm3.231.3±2.631.3\pm2.69.6±2.19.6\pm2.182.0±2.782.0\pm2.725.3±2.625.3\pm2.684.8±1.684.8\pm1.633.3±3.833.3\pm3.8
ThoughtComm78.5±3.078.5\pm3.039.6±3.239.6\pm3.228.3±2.628.3\pm2.616.7±2.616.7\pm2.665.0±3.365.0\pm3.344.0±2.944.0\pm2.970.0±2.170.0\pm2.142.0±4.042.0\pm4.0
Interlat60.0±3.560.0\pm3.528.0±2.728.0\pm2.712.7±1.912.7\pm1.920.7±2.820.7\pm2.872.5±3.272.5\pm3.231.3±2.731.3\pm2.764.0±2.464.0\pm2.438.0±4.038.0\pm4.0
SwiReasoning81.5±2.681.5\pm2.673.2±2.773.2\pm2.730.3±2.730.3\pm2.715.7±2.615.7\pm2.681.5±2.881.5\pm2.851.3±2.851.3\pm2.882.0±1.882.0\pm1.863.3±4.163.3\pm4.1
Soft Thinking84.5±2.584.5\pm2.567.6±3.067.6\pm3.033.7±2.833.7\pm2.815.7±2.615.7\pm2.683.0±2.683.0\pm2.649.0±2.849.0\pm2.881.0±1.981.0\pm1.960.0±4.060.0\pm4.0
SoftCoT53.0±3.553.0\pm3.544.4±3.144.4\pm3.136.3±2.836.3\pm2.823.7±3.023.7\pm3.057.5±3.557.5\pm3.546.7±2.946.7\pm2.954.2±2.454.2\pm2.426.7±3.626.7\pm3.6
Ours89.0±2.289.0\pm2.266.0±3.066.0\pm3.047.0±3.047.0\pm3.031.3±3.331.3\pm3.384.5±2.584.5\pm2.554.7±2.854.7\pm2.883.0±1.883.0\pm1.833.3±3.833.3\pm3.8

Table 1. Accuracy (%) against baselines. Bold and underlined values indicate the best and second-best among latent methods.

Observation 2. Latent belief interaction remains effective across model scales.

Scaling with model size. Table 2 and Figure 4 demonstrate that SoB maintains competitive performance across the Qwen3 and Llama families without changing its inference configuration. While text-based debate retains advantages on certain tasks, the consistent effectiveness of SoB across model scales suggests that latent collaboration can benefit from increasingly capable base models.

Scaling with model size across MATH-500, GPQA-D, MedQA, and BBH for the Qwen3 and Llama families

Figure 4. Scaling with model size. Accuracy across benchmarks for the Qwen3 family (top) and Llama family (bottom).

Figure 5. (a) Problem #18 from graph connectivity. (b) Problem #7 from shortest path. (c) Problem #13 from ProsQA. Qualitative demonstration of latent belief trajectories. For each problem, the latent beliefs of N=4N = 4 agents (Qwen3–8B, L=48L = 48) over cycles k=0,…,16k = 0, \ldots, 16 are projected onto 3D with PCA. Lines show the collective trajectory (mean over agents), arrows indicate the progression from one cycle to the next, and the shaded region spans the agents’ beliefs at consecutive cycles. Blue and red denote that the collective answer is correct and incorrect, respectively. No communication: agents refine their beliefs in isolation; Belief iteration: agents interact through their latent beliefs; Forced consensus: all beliefs are replaced by their mean after every cycle. The right column shows the belief residual rkr_k; the vertical line marks the equilibrium cycle k^\hat{k} at which rk<ϵr_k < \epsilon. Examples are from distributed-information tasks in which each agent observes only part of the problem.

Efficiency

Observation 3. Parallel latent belief interaction enables efficient collective reasoning with reduced sequential latency and explicit text generation.

Inference Latency. We measure token latency by sequential generation depth, including both latent and text tokens. As shown in Figure 6, SoB achieves shorter sequential latency than LatentMAS, the fastest baseline, reducing token latency by 1.1×1.1\times–2.6×2.6\times across the eight benchmarks (1.5×1.5\times on average) while maintaining competitive reasoning accuracy across different belief lengths. This efficiency reflects a different organization of collective reasoning: rather than extending sequential reasoning chains, SoB distributes computation across parallel belief trajectories.

Latency and token consumption plots

Figure 6. Latency and token consumption. (Qwen3-8B). Top left: token latency, the number of tokens generated one after another before the answer. Bottom left: text tokens per question, summed over all agents. Right: accuracy against token latency; our method for belief lengths L∈{96,192,384}L \in\{96, 192, 384\} (ringed: L=384L = 384), SwiReasoning and Soft Thinking with budgets from 512 to 8,192 tokens, and LatentMAS with judger caps of 256, 512, and 1,024 tokens.

Text-Token Usage. Figure 6 further shows that SoB substantially reduces explicit text generation compared with Soft Thinking, the baseline with the lowest text-token usage: 51% fewer text tokens on average, and up to 82% fewer on individual benchmarks. By exchanging continuous beliefs instead of textual reasoning traces, SoB decouples inter-agent communication from language generation, reserving text primarily for intermediate probes and final readouts. These gains reflect reduced sequential depth and textual communication, rather than a reduction in total computational cost.

Distributions of Latent Representations

Observation 4. Continuous belief tokens spontaneously align with vocabulary embeddings through collective belief dynamics.

Spontaneous alignment. By construction, belief tokens are probability-weighted mixtures of vocabulary embeddings ((Eq. 1)), allowing multiple candidate tokens to contribute to subsequent computation. Reremarkably, as collective beliefs evolve, these continuous representations spontaneously approach individual vocabulary embeddings (Figure 8). Their mean cosine distance to the nearest vocabulary embedding decreases from 0.020 at k=0k = 0 to approximately 3×10−43 \times10^{-4} at k=16k = 16, with a characteristic timescale of τ≈1.9\tau\approx1.9 cycles. This spontaneous alignment suggests that token-level commitment can emerge naturally from continuous belief dynamics, without explicit discretization.

Figure 8. Quantitative demonstration of spontaneous sharpening of belief tokens. (a) Energy distance between the latent inputs used by each method and the input embeddings of the tokens generated by the same model in its text chain of thought on the same questions. (b) Averaged distance (1 − cos) between a belief token and its nearest vocabulary embedding across cycles of belief iteration (Qwen3-8B, N=4N = 4, L=192L = 192, BBH logical-deduction questions). Belief tokens start as genuine mixtures of multiple token embeddings and rapidly approach individual vocabulary embeddings over successive cycles.

Distributional compatibility. To assess how closely latent inputs resemble the model’s native token representations, we measure their energy distance (ED) [69] to embeddings of tokens generated by the same model on the same questions:

ED⁡(X,Y)=2E∥X−Y∥−E∥X−X′∥−E∥Y−Y′∥,(6)\operatorname{ED}(X,Y) = 2\mathbb{E}\lVert X-Y\rVert-\mathbb{E}\lVert X-X'\rVert-\mathbb{E}\lVert Y-Y'\rVert, \tag*{(6)}

where XX, X′X' and YY, Y′Y' are independent samples from the latent-input and text-token embedding distributions, respectively. We estimate these expectations using 4,000 vectors per distribution in the full embedding space. Belief tokens remain close to the model’s native text-token distribution throughout iteration (Figure 7), exhibiting substantially smaller distributional shifts than alternative latent reasoning methods (Figure 8a).

Figure 7. Latent representation distributions vs text-token distribution. Top: Spontaneous alignment of the distribution of our constructed belief tokens. Bottom: the distributions of latent representations from baseline methods. Corner of each panel: energy distance (ED) to the text-token distribution in the full representation space and mean cosine between a vector and its nearest vocabulary embedding (Qwen3-8B, BBH logical deduction).

Ablations

Observation 5. The capacity to represent beliefs matters more than the extent of collective deliberation.

Quality Factors. Figure 9 examines how belief capacity, agent population, interaction cycles, and local-refinement depth shape collective reasoning. Among these factors, belief capacity exhibits the strongest and most consistent influence, suggesting that richer latent representations provide a more effective foundation for collective reasoning. In contrast, increasing the number of agents yields task-dependent benefits, indicating that a larger population does not necessarily translate into stronger collective intelligence.

Figure 9. Ablations of the hyperparameters of SoB (Qwen3-8B). Boxes: 68% (dark) and 95% (light) bootstrap intervals.

Extending deliberation offers similarly limited benefits. Increasing the number of interaction cycles or local refinements produces no consistent improvement in accuracy, despite requiring additional belief updates. Together, these findings suggest that effective collective reasoning depends less on repeatedly revisiting beliefs than on the representational capacity available to each agent.

Related Work

Multi-agent reasoning with language models. Recent work extends multi-agent systems to LLM-based agents [36, 41, 56, 72, 82] that collaborate on reasoning, planning, and problem-solving [71, 76, 93]. Existing methods coordinate agents through dialogue and role assignment [48, 79, 85], structured communication [86], and emergent specialization [44, 55]. Multi-agent debate improves reasoning and factuality through iterative proposal and critique [11, 24, 52], while general frameworks organize agents into roles and workflows [13, 41, 48, 56, 79]. Other work optimizes interaction graphs [97], generates task-specific topologies [89], and studies scaling through populations [49, 59]. Our work formulates multi-agent reasoning as the evolution of a joint latent belief state toward a collective equilibrium.

Latent interaction between agents. Recent MAS have also explored interaction through latent representations [18, 28, 72, 86, 94]. Existing work studied intermediate activations [60], last-layer hidden states [25], hidden-state deltas [70], or key–value caches [28, 64, 94, 98]. We introduce belief tokens, a belief iteration operator, and a belief residual to characterize the latent states of reasoning agents.

Test-time scaling of language model reasoning. Chain-of-thought [77], self-consistency [74], and tree search [84] scale test-time computation through explicit text, and allocating that computation adaptively can outperform scaling model size [65]. Latent reasoning moves this computation into continuous space [96]. Training-based approaches add pause tokens [34], distill explicit chains into hidden states [22, 62], feed hidden states back as continuous thoughts [37], generate soft thoughts with an assistant model [80], iterate a recurrent block in depth [30], or scale and optimize latent thoughts at test time [23, 88]. Training-free approaches replace sampled tokens with mixtures of embeddings [92] or switch between latent and explicit thinking according to the model’s confidence [63]. Our method organizes inference as alternating global and local belief updates and stops when the collective state stabilizes, shifting the focus from scaling computation to shaping its dynamics.

Belief and opinion dynamics in collective computation. Classical models of social influence study how locally interacting opinions evolve toward consensus, polarization, clustering, or persistent disagreement, depending on the update rule and interaction graph [10, 20, 21, 27, 38], and when such interactions aggregate dispersed information or amplify errors [3, 32, 46]. Related ideas appear in distributed computation, where belief propagation, consensus protocols, and collective dynamical systems compute global quantities through repeated local updates and relaxation toward stable states [9, 42, 57, 73, 87]. Recent work has begun to observe analogous phenomena in populations of language-model agents interacting through text, including consensus, conformity, shared conventions, and collective biases [6, 14, 15, 26, 90]. We extend these questions to agents whose states are high-dimensional latent beliefs updated by a frozen language model through nonlinear dynamics. Instead of assuming classical convergence guarantees, we empirically study whether these beliefs settle, how interaction reshapes the collective state, and whether equilibrium requires agreement.

Conclusion

We introduced Stream of Belief (SoB), formulating multi-agent reasoning as the evolution of interacting latent beliefs toward a stable joint state. By coupling global interaction with local refinement, SoB enables agents to collaborate through high-dimensional semantic spaces via belief tokens. Across diverse reasoning tasks and model families, our experiments demonstrate the effectiveness and efficiency of this approach, while revealing spontaneous alignment between evolving beliefs and the model’s native token representations. More broadly, our findings suggest that collective intelligence can emerge through latent belief dynamics, offering a complementary perspective on multi-agent reasoning beyond language-based communication.

Broad Impact. SoB opens a direction for studying collective intelligence as a dynamical process in continuous representation spaces. This perspective may inform more efficient and expressive multi-agent systems, while motivating further investigation into how information propagates, beliefs interact, and collective behavior emerges in language models.

Limitations. Our framework relies on frozen language models and fixed-length belief representations, which may constrain the complexity of information exchanged among agents. Moreover, stable decoded predictions do not guarantee convergence to a latent fixed point or correctness of the collective solution. Finally, while latent communication reduces sequential generation and explicit text usage, it introduces computational overhead and makes intermediate reasoning less directly interpretable than natural-language communication.

Acknowledgments

This work is supported by the Defense Advanced Research Projects Agency (DARPA) under the Young Faculty Award (YFA) program. The views and conclusions are those of the authors and should not be interpreted as representing the official policies of DARPA or the U.S. Government.

Appendices

Prompt Templates of Stream of Belief

Every belief update and every decoding step is one chat-formatted call of the frozen model (system turn, user turn, assistant turn in the model’s own chat template), identical for all models and benchmarks. ⟨zi⟩\langle z_i\rangle denotes the belief of agent ii, inserted as LL soft input embeddings; {problem} is the problem text, with multiple-choice options and listed facts shown in an agent-specific random order.

System turn (all calls)

You are Agent ii in a team of NN agents solving a reasoning problem together. {persona}

Personas (agent ii receives persona i mod 8i \bmod8):

  1. You reason carefully and check each step.

  2. You look for the most direct chain of evidence.

  3. You are skeptical and try to rule out wrong options.

  4. You organise the facts systematically before concluding.

  5. You focus on tracing relationships one hop at a time.

  6. You look for contradictions and edge cases.

  7. You summarise what is certain before deciding.

  8. You verify the answer by working backwards from the options.

Initial belief (k=0k = 0): zi←Ai(x)z_i \leftarrow A_i(x)

User: {problem}

Concisely work through the key reasoning steps and end with your conclusion.

Assistant: ⟨zi⟩\langle z_i\rangle (LL belief tokens, Eq. (1))

Global interaction: [G(Z)]i=Ai(x,zi,{zj}j≠i)[G(Z)]_i = A_i(x, z_i, \{z_j\}_{j\ne i})

User: {problem}

Conclusions shared by other agents (in their own latent words):

[Agent jj]: ⟨zj⟩\langle z_j\rangle (one line per other agent j≠ij \ne i)

Your previous conclusion: ⟨zi⟩\langle z_i\rangle

Reconsider the problem and the conclusions above. Concisely work through the key reasoning steps and end with your revised conclusion.

Assistant: ⟨zi⟩\langle z_i\rangle (new belief, replaces ⟨zi⟩\langle z_i\rangle)

Local refinement: [L(Z)]i=Ai(x,zi)[L(Z)]_i = A_i(x, z_i)

User: {problem}

Your previous conclusion: ⟨zi⟩\langle z_i\rangle

Reconsider the problem and your previous conclusion. Concisely work through the key reasoning steps and end with your revised conclusion.

Assistant: ⟨zi⟩\langle z_i\rangle (new belief, replaces ⟨zi⟩\langle z_i\rangle)

Probe (after each cycle; 24 text tokens)

User: {problem} Work through the problem, then give the final answer. Assistant: ⟨zi⟩\langle z_i\rangle Therefore, the final answer is: Answer: {decoded answer}

Final solution (after stopping; up to 512 text tokens)

User: {problem} Work through the problem, then end with ‘Answer: <answer>’. Assistant: ⟨zi⟩\langle z_i\rangle Final solution (brief): {decoded solution, ending with “Answer: <answer>”}

In the probe and the final solution, the assistant turn begins with the agent’s own belief ⟨zi⟩\langle z_i\rangle, followed by a text cue after which the model writes text; the decoder sees only the agent’s own belief, not those of the other agents. The probe is used by the stopping rule only (Algorithm 1); the final prediction is the majority over the agents’ final solutions.

Baseline Implementation Details

Multi-agent debate. We follow the official implementation of Du et al. [24] in its default configuration (three agents, two rounds, where the first round is the independent answer): three agents of the same base model receive the question with the official prompt (“Can you solve the following math problem? . . . ” for mathematics, “Can you answer the following question as accurately as possible? . . . ” otherwise, with the answer requested as \boxed{}, “(X)”, “(yes)/(no)” or “Answer: . . . ” according to the benchmark), and in the debate round every agent is shown the full solutions of the other two agents and asked for an updated answer with the official debate prompt.

LatentMAS. We use the official implementation with the sequential-agent configuration, 40 latent steps per agent, temperature 0.6, and top-pp 0.95.

ThoughtComm. We use the official implementation with three agents and two communication rounds of 256 tokens. For each base model, the autoencoder, communication structure, and prefix adapter are trained on 300 held-out problems following the released training schedule.

Interlat. We follow the official sender–receiver pipeline. Sender plans are collected from 4,000 MATH training problems. The receiver is trained with LoRA of rank 64 on all linear layers together with trainable input embeddings, while the hidden-state adapter is trained for 3 epochs with learning rate 10−410^{-4}. At inference time, sender hidden states are inserted between the plan delimiters and the receiver decodes greedily.

SoftCoT. We use the official implementation with 32 soft-thought tokens and train the projection on GSM8K for 3 epochs. Decoding uses the default temperature of 1.0, with a generation budget of 4,096 tokens for MATH-500 and AIME and 2,048 tokens otherwise.

SwiReasoning. We use the official implementation with the dataset-specific hyperparameters released by the authors. The sequential decoding budget is capped at the mean number of sequential steps used by our method on the same benchmark and base model.

Soft Thinking. We implement the method in HuggingFace Transformers following the official algorithm and defaults: at most 10 concept tokens per step, top-pp 0.95, top-kk 30, min-pp 0.001, temperature 0.6, and cold stopping after 256 steps with entropy below 0.01. The sequential decoding budget is matched to our method on each benchmark and base model.

References

  1. [1]R. P. Abelson. Mathematical models of the distribution of attitudes under controversy. In N. Fred-eriksen and H. Gulliksen, editors, Contributions to Mathematical Psychology, pages 142–160. Holt, Rinehart and Winston, New York, 1964.
  2. [2]A. Abouelenin, A. Ashfaq, A. Atkinson, H. Awadalla, N. Bach, J. Bao, et al. Phi-4-Mini technical report: Compact yet powerful multimodal language models via mixture-of-LoRAs, 2025.arxiv.org/abs/2503.01743
  3. [3]D. Acemoglu and A. Ozdaglar. Opinion dynamics and learning in social networks. Dynamic Games and Applications, 1(1):3–49, 2011. doi: 10.1007/s13235-010-0004-1.DOI
  4. [4]D. Acemoglu, G. Como, F. Fagnani, and A. Ozdaglar. Opinion fluctuations and disagreement in social networks. Mathematics of Operations Research, 38(1):1–27, 2013. doi: 10.1287/moor.1120.0570.DOI
  5. [5]D. B. Acharya, K. Kuppan, and B. Divya. Agentic AI: Autonomous intelligence for complex goals—a comprehensive survey. IEEE Access, 13:18912–18936, 2025. doi: 10.1109/ACCESS.2025.3532853.DOI
  6. [6]A. F. Ashery, L. M. Aiello, and A. Baronchelli. Emergent social conventions and collective bias in LLM populations. Science Advances, 11(20):eadu9368, 2025. doi: 10.1126/sciadv.adu9368.DOI
  7. [7]J. Becker, D. Brackbill, and D. Centola. Network dynamics of social influence in the wisdom of crowds. Proceedings of the National Academy of Sciences, 114(26):E5070–E5076, 2017. doi: 10.1073/pnas.1615978114.DOI
  8. [8]D. Bindel, J. Kleinberg, and S. Oren. How bad is forming your own opinion? Games and Economic Behavior, 92:248–265, 2015. doi: 10.1016/j.geb.2014.06.004.DOI
  9. [9]S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah. Randomized gossip algorithms. IEEE Transactions on Information Theory, 52(6):2508–2530, 2006. doi: 10.1109/TIT.2006.874516.DOI
  10. [10]G. Castellano, S. Fortunato, and V. Loreto. Statistical physics of social dynamics. Reviews of Modern Physics, 81(2):591–646, 2009. doi: 10.1103/RevModPhys.81.591.DOI
  11. [11]C.-M. Chan, W. Chen, Y. Su, J. Yu, W. Xue, S. Zhang, J. Fu, and Z. Liu. ChatEval: Towards better LLM-based evaluators through multi-agent debate. In International Conference on Learning Representations, 2024.arxiv.org/abs/2308.07201
  12. [12]J. Chen, S. Saha, and M. Bansal. Reconcile: Round-table conference improves reasoning via consensus among diverse LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7066–7085, 2024. doi: 10.18653/v1/2024.acl-long.381.DOI
  13. [13]W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C.-M. Chan, H. Yu, Y. Lu, Y.-H. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learning Representations, 2024.arxiv.org/abs/2308.10848
  14. [14]Y.-S. Chuang, A. Goyal, N. Harlalka, S. Suresh, R. Hawkins, S. Yang, D. Shah, J. Hu, and T. T. Rogers. Simulating opinion dynamics with networks of LLM-based agents. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3326–3346, 2024. doi: 10.18653/v1/2024.findings-naacl.211.
  15. [15]P. Cisneros-Velarde. Biases in opinion dynamics in multi-agent systems of large language models: A case study on funding allocation. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 1889–1916, 2025. doi: 10.18653/v1/2025.findings-naacl.101.
  16. [16]P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord. Think you have solved question answering? try ARC, the AI2 reasoning challenge, 2018.arxiv.org/abs/1803.05457
  17. [17]K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, G. Hesse, and J. Schulman. Training verifiers to solve math word problems, 2021.arxiv.org/abs/2110.14168
  18. [18]J. Goda-Formo, Z. Zhao, Q. Zhang, D. Tamboli, W. Li, X. Fan, L. Zhang, E. Schulz, and H.-P. Tseng. Exploring system 1 and 2 communication for latent reasoning in llms. arXiv preprint arXiv:2510.00494, 2025.
  19. [19]I. D. Couzin. Collective cognition in animal groups. Trends in Cognitive Sciences, 13(1):36–43, 2009. doi: 10.1016/j.tics.2008.10.002.DOI
  20. [20]G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch. Mixing beliefs among interacting agents. Advances in Complex Systems, 3(1–4):87–98, 2000. doi: 10.1142/S0219525900000078.DOI
  21. [21]M. H. DeGroot. Reaching a consensus. Journal of the American Statistical Association, 69(345):118–121, 1974. doi: 10.1080/01621459.1974.10480137.DOI
  22. [22]Y. Deng, K. Prasad, R. Fernandez, P. Smolensky, V. Chaudhary, and S. Shieber. Implicit chain of thought reasoning via knowledge distillation, 2023.arxiv.org/abs/2311.01460
  23. [23]H. Du, Y. Dong, and X. Ning. Latent thinking optimization: Your latent reasoning language model secretly encodes reward signals in its latent thoughts. In International Conference on Learning Representations, 2026.arxiv.org/abs/2509.26314
  24. [24]Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 11733–11763, 2024.arxiv.org/abs/2305.14325
  25. [25]Z. Du, R. Wang, H. Bai, Z. Cao, X. Zhu, Y. Cheng, B. Zheng, W. Chen, and H. Ying. Enabling agents to communicate entirely in latent space. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 27106–27129. Association for Computational Linguistics, 2026. doi: 10.18653/v1/2026.acl-long.1248.
  26. [26]A. Estornell and Y. Liu. Multi-LLM debate: Framework, principals, and interventions. In Advances in Neural Information Processing Systems, volume 37, pages 28938–28964, 2024.DOI
  27. [27]N. E. Friedkin and E. C. Johnsen. Social influence and opinions. The Journal of Mathematical Sociology, 15(3-4):193–206, 1990. doi: 10.1080/0022250X.1990.9990069.DOI
  28. [28]T. Fu, Z. Min, H. Zhang, J. Yan, G. Dai, W. Ouyang, and Y. Wang. Cache-to-cache: Direct semantic communication between large language models. In International Conference on Learning Representations, 2026.arxiv.org/abs/2510.03215
  29. [29]F. Galton. Vox populi. Nature, 75(1949):450–451, 1907. doi: 10.1038/075450a0.DOI
  30. [30]J. Geiping, S. McLeish, N. Jain, J. Kirchenbauer, S. Singh, B. R. Bartoldson, B. Kailkhura, A. Bhatele, and T. Goldstein. Scaling up test-time compute with latent reasoning: A recurrent depth approach. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2502.05171
  31. [31]Gemma Team. Gemma 3 technical report, 2025.
  32. [32]B. Golub and M. O. Jackson. Naive learning in social networks and the wisdom of crowds. American Economic Journal: Microeconomics, 2(1):112–149, 2010. doi: 10.1257/mic.2.1.112.DOI
  33. [33]J. Gottweis, W.-H. Weng, A. Daryin, T. Tu, P. Sirkovic, A. Myaskovsky, G. Glowaty, F. Weissenberger, A. Orlandi, D. Popovici, A. Palepu, K. Rong, et al. Accelerating scientific discovery with Co-Scientist. Nature, 655(8122):487–496, 2026. doi: 10.1038/s41586-026-10644-y.DOI
  34. [34]S. Goyal, Z. Ji, A. S. Rawat, A. K. Menon, S. Kumar, and V. Nagarajan. Think before you speak: Training language models with pause tokens. In International Conference on Learning Representations, 2024.arxiv.org/abs/2310.02226
  35. [35]A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, et al. The Llama 3 herd of models, 2024.arxiv.org/abs/2407.21783
  36. [36]T. Guo, X. Chen, Y. Wang, R. Chang, S. Pei, N. V. Chawla, O. Wiest, and X. Zhang. Large language model based multi-agents: A survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24), pages 8048–8057, 2024. doi: 10.24963/ijcai.2024/890.
  37. [37]S. Hao, S. Sukhbaatar, D. Su, X. Li, Z. Hu, J. Weston, and Y. Tian. Training large language models to reason in a continuous latent space. In Conference on Language Modeling, 2025.arxiv.org/abs/2412.06769
  38. [38]R. Hegselmann and U. Krause. Opinion dynamics and bounded confidence: Models, analysis and simulation. Journal of Artificial Societies and Social Simulation, 5(3):2, 2002.
  39. [39]D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Steinhardt. Measuring mathematical problem solving with the MATH dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1, 2021.
  40. [40]L. Hong and S. E. Page. Groups of diverse problem solvers can outperform groups of high-ability problem solvers. Proceedings of the National Academy of Sciences, 101(46):16385–16389, 2004. doi: 10.1073/pnas.0403723101.DOI
  41. [41]S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, L. Zhou, C. Ran, L. Xiao, C. Wu, and J. Schmidhuber. MetaGPT: Meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, 2024.arxiv.org/abs/2308.00352
  42. [42]J. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8):2554–2558, 1982. doi: 10.1073/pnas.79.8.2554.DOI
  43. [43]M. Hu, Y. Zhou, W. Fan, Y. Nie, B. Xia, T. Sun, Z. Ye, Z. Jin, Y. Li, Q. Chen, et al. OWL: Optimized workforce learning for general multi-agent assistance in real-world task automation. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2505.23885
  44. [44]Q. Huang, Z. Zhou, Y. Li, K. Yang, B. Wang, and Y. Wang. Many minds, one goal: Time series forecasting via sub-task specialization and inter-agent cooperation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URL https://openreview.net/forum?id=Uon41HfqR3.DOI
  45. [45]M. A. Islam, M. E. Ali, and M. R. Parvez. MapCoder: Multi-agent code generation for competitive problem solving. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4912–4944, 2024. doi: 10.18653/v1/2024.acl-long.269.DOI
  46. [46]J. Jadbabaie, P. Molavi, A. Sandroni, and A. Tahbaz-Salehi. Non-Bayesian social learning. Games and Economic Behavior, 76(1):210–225, 2012. doi: 10.1016/j.geb.2012.06.001.DOI
  47. [47]D. Jin, E. Pan, N. Oufattolle, W.-H. Weng, H. Fang, and P. Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 11(14):6421, 2021. doi: 10.3390/app11146421.
  48. [48]G. Li, H. A. A. K. Hammoud, H. Itani, D. Khizbullin, and B. Ghanem. CAMEL: Communicative agents for “mind” exploration of large language model society. In Advances in Neural Information Processing Systems, volume 36, 2023.arxiv.org/abs/2303.17760
  49. [49]J. Li, Q. Zhang, Y. Yu, Q. Fu, and D. Ye. More agents is all you need. Transactions on Machine Learning Research, 2024.arxiv.org/abs/2402.05120
  50. [50]Z. Li, H. Zhang, S. Han, S. Liu, J. Xie, Y. Zhang, Y. Choi, J. Zou, and P. Lu. In-the-flow agentic system optimization for effective planning and tool use. arXiv preprint arXiv:2510.05592, 2025.arxiv.org/abs/2510.05592
  51. [51]Z. Li, J. Zou, F. Fang, X. Ning, M. Ai, T. Wei, S. Chen, X. Yang, and J. He. Heterogeneous scientific foundation model collaboration. CoRR, abs/2604.27351, 2026. doi: 10.48550/ARXIV.2604.27351. URL https://doi.org/10.48550/arXiv.2604.27351.
  52. [52]T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, and Z. Tu. Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17889–17904, 2024.arxiv.org/abs/2305.19118
  53. [53]H. Lightman, V. Kosaraju, Y. Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Gobbe. Let’s verify step by step. In International Conference on Learning Representations, 2024.arxiv.org/abs/2305.20050
  54. [54]J. Lorenz, H. Rauhut, F. Schweitzer, and D. Helbing. How social influence can undermine the wisdom of crowd effect. Proceedings of the National Academy of Sciences, 108(22):9020–9025, 2011. doi: 10.1073/pnas.1008636108.DOI
  55. [55]E. Mieczkowski, R. Mon-Williams, N. Bramley, C. G. Lucas, N. Velez, and T. L. Griffiths. Predicting multi-agent specialization via task parallelizability. arXiv preprint arXiv:2503.15703, 2025.arxiv.org/abs/2503.15703
  56. [56]J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023. doi: 10.1145/3586183.3606763.DOI
  57. [57]J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, San Mateo, CA, 1988.
  58. [58]C. Qian, W. Liu, H. Liu, N. Chen, Y. Dang, J. Li, C. Yang, W. Chen, Y. Su, X. Cong, J. Xu, D. Li, Z. Liu, and M. Sun. ChatDev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174–15186, 2024. doi: 10.18653/v1/2024.acl-long.810.
  59. [59]C. Qian, Z. Xie, Y. Wang, W. Liu, K. Zhu, H. Xia, Y. Dang, Z. Du, W. Chen, C. Yang, Z. Liu, and M. Sun. Scaling large language model-based multi-agent collaboration. In International Conference on Learning Representations, 2025.arxiv.org/abs/2406.07155
  60. [60]V. Ramesh and K. Li. Communicating activations between language model agents. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 51094–51116. PMLR, 2025.arxiv.org/abs/2501.14082
  61. [61]D. Rein, B. L. Hou, A. C. Stickland, J. Petty, R. Y. Pang, J. Dirani, J. Michael, and S. R. Bowman. GPQA: A graduate-level Google-proof Q&A benchmark. In Conference on Language Modeling, 2024.arxiv.org/abs/2311.12022
  62. [62]Z. Shen, H. Yan, L. Zhang, Z. Hu, Y. Du, and Y. He. CODI: Compressing chain-of-thought into continuous space via self-distillation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 677–693. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main.36.
  63. [63]D. Shi, A. Asi, K. Li, X. Yuan, L. Pan, W. Lee, and W. Xiao. SwiReasoning: Switch-thinking in latent and explicit for pareto-superior reasoning LLMs. In International Conference on Learning Representations, 2026.arxiv.org/abs/2510.05069
  64. [64]X. Shi, M. Chiesa, G. Q. Maguire Jr., and D. Kostic. KVComm: Enabling efficient LLM communication through selective KV sharing. In International Conference on Learning Representations, 2026.arxiv.org/abs/2510.03346
  65. [65]C. Snell, J. Lee, K. Xu, and A. Kumar. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In International Conference on Learning Representations, 2025.arxiv.org/abs/2408.03314
  66. [66]J. Surowiecki. The Wisdom of Crowds: Why the Many Are Smarter Than the Few and How Collective Wisdom Shapes Business, Economies, Societies and Nations. Doubleday, New York, 2004.DOI
  67. [67]M. Suzgun, N. Scales, N. Schärli, S. Gehrmann, Y. Tay, H. W. Chung, A. Chowdhery, Q. Le, E. Chi, D. Zhou, and J. Wei. Challenging BIG-bench tasks and whether chain-of-thought can solve them. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13003–13051. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.findings-acl.824.DOI
  68. [68]K. Swanson, W. Wu, N. L. Bulaong, J. E. Pak, and J. Zou. The virtual lab of AI agents designs new SARS-CoV-2 nanobodies. Nature, 646(8085):716–723, 2025. doi: 10.1038/s41586-025-09442-9.DOI
  69. [69]G. J. Székely and M. L. Rizzo. Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143(8):1249–1272, 2013. doi: 10.1016/j.jspi.2013.03.018.DOI
  70. [70]Y. Tang, W. Su, Y. Zhou, Y. Liu, M. Zhang, S. Ma, and Q. Ai. Augmenting multi-agent communication with state delta trajectory. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10219–10240. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main.518.DOI
  71. [71]W. Tao, Y. Zhou, Y. Wang, W. Zhang, H. Zhang, and Y. Cheng. MAGIS: LLM-based multi-agent framework for GitHub issue resolution. In Advances in Neural Information Processing Systems, volume 37, pages 51963–51993, 2024.arxiv.org/abs/2403.17927
  72. [72]K.-T. Tran, D. Dao, M.-D. Nguyen, Q.-V. Pham, B. O’Sullivan, and H. D. Nguyen. Multi-agent collaboration mechanisms: A survey of llms. arXiv preprint arXiv:2501.06322, 2025.arxiv.org/abs/2501.06322
  73. [73]J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans. Distributed asynchronous deterministic and stochastic gradient optimization algorithms. IEEE Transactions on Automatic Control, 31(9):803–812, 1986. doi: 10.1109/TAC.1986.1104412.DOI
  74. [74]X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations, 2023.arxiv.org/abs/2203.11171
  75. [75]Y. Wang, X. Ma, G. Zhang, Y. Ni, A. Chandra, S. Guo, W. Ren, A. Arulraj, X. He, Z. Jiang, T. Li, M. Ku, K. Wang, A. Zhuang, R. Fan, X. Yue, and W. Chen. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2406.01574
  76. [76]Z. Wang, S. Moriyama, W.-Y. Wang, B. Gangopadhyay, and S. Takamatsu. Talk structurally, act hierarchically: A collaborative framework for LLM multi-agent systems. IEEE Transactions on Artificial Intelligence, 7(9):4993–5003, 2026. doi: 10.1109/TAI.2026.3676025.DOI
  77. [77]J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, 2022.arxiv.org/abs/2201.11903
  78. [78]A. W. Woolley, C. F. Chabris, A. Pentland, N. Hashmi, and T. W. Malone. Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004):686–688, 2010. doi: 10.1126/science.1193147.
  79. [79]Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversation. In Conference on Language Modeling, 2024.arxiv.org/abs/2308.08155
  80. [80]Y. Xu, X. Guo, Z. Zeng, and C. Miao. SoftCoT: Soft chain-of-thought for efficient reasoning with LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 23336–23351. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.acl-long.1137.DOI
  81. [81]A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.arxiv.org/abs/2505.09388
  82. [82]Y. Yang, Q. Peng, J. Wang, Y. Wen, and W. Zhang. Llm-based multi-agent systems: Techniques and business perspectives. arXiv preprint arXiv:2411.14033, 2024.arxiv.org/abs/2411.14033
  83. [83]Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369–2380. Association for Computational Linguistics, 2018. doi: 10.18653/v1/D18-1259.DOI
  84. [84]S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y. Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, volume 36, pages 11809–11822, 2023.arxiv.org/abs/2305.10601
  85. [85]S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations, 2023.arxiv.org/abs/2210.03629
  86. [86]H. Ye, Z. Gao, M. Ma, Q. Wang, Y. Fu, M.-Y. Chung, Y. Lin, Z. Liu, J. Zhang, D. Zhuo, et al. KVCOMM: Online cross-context KV-cache communication for efficient LLM-based multi-agent systems. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2510.12872
  87. [87]J. S. Yedidia, W. T. Freeman, and Y. Weiss. Constructing free-energy approximations and generalized belief propagation algorithms. IEEE Transactions on Information Theory, 51(7):2282–2312, 2005. doi: 10.1109/TIT.2005.850085.DOI
  88. [88]R. You, Y. Li, M. Liu, W. Wang, L. Nie, and W. Li. Parallel test-time scaling for latent reasoning models. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 44703–44717. Association for Computational Linguistics, 2026. doi: 10.18653/v1/2026.acl-long.2069.DOI
  89. [89]G. Zhang, Y. Yue, X. Sun, G. Wan, M. Yu, J. Fang, K. Wang, T. Chen, and D. Cheng. G-Designer: Architecting multi-agent communication topologies via graph neural networks. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 76678–76692, 2025.arxiv.org/abs/2410.11782
  90. [90]J. Zhang, X. Xu, N. Zhang, R. Liu, B. Hooi, and S. Deng. Exploring collaboration mechanisms for LLM agents: A social psychology view. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14544–14607, 2024. doi: 10.18653/v1/2024.acl-long.782.DOI
  91. [91]Y. Zhang, R. Sun, Y. Chen, T. Pfister, R. Zhang, and S. Arik. Chain of agents: Large language models collaborating on long-context tasks. In Advances in Neural Information Processing Systems, volume 37, pages 132208–132237, 2024.arxiv.org/abs/2406.02818
  92. [92]Z. Zhang, X. He, W. Yan, A. Shen, C. Zhao, and X. E. Wang. Soft thinking: Unlocking the reasoning potential of LLMs in continuous concept space. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2505.15778
  93. [93]W. Zhao, M. Yuksekgonul, S. Wu, and J. Zou. Sirius: Self-improving multi-agent systems via bootstrapped reasoning. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2502.04780
  94. [94]Y. Zheng, Z. Zhao, Z. Li, Y. Xie, M. Gao, L. Zhang, and K. Zhang. Thought communication in multiagent collaboration. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2510.20733
  95. [95]H. Zhou, H. Geng, X. Xue, L. Kang, Y. Qin, Z. Wang, Z. Yin, and L. Bai. ReSo: A reward-driven self-organizing LLM-based multi-agent system for reasoning tasks. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 15979–15998, 2025. doi: 10.18653/v1/2025.emnlp-main.808.DOI
  96. [96]R.-J. Zhu, T. Peng, T. Cheng, X. Qu, J. Huang, D. Zhu, H. Wang, K. Xue, X. Zhang, Y. Shan, T. Cai, T. Kergan, A. Kembay, A. Smith, C. Lin, et al. A survey on latent reasoning, 2025.arxiv.org/abs/2507.06203
  97. [97]M. Zhuge, W. Wang, L. Kirsch, F. Faccio, D. Khizbullin, and J. Schmidhuber. GPTSwarm: Language agents as optimizable graphs. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 62743–62767, 2024.arxiv.org/abs/2402.16823
  98. [98]J. Zou, R. Qiu, G. Li, X. Yang, K. Tieu, P. Lu, K. Shen, H. Tong, Y. Choi, J. He, J. Zou, M. Wang, and L. Yang. Latent collaboration in multi-agent systems. In International Conference on Machine Learning, 2026.arxiv.org/abs/2511.20639

Paper details

vlnc:W_txgvcbga
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents