Skip to content

Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; visual MAS hallucination snowballing — seeded in one agent, amplified across

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Visual Language Model (VLM)-based Multi-Agent Systems (MAS) suffer from "hallucination snowballing," where early visual misinterpretations are amplified by downstream agents over-relying on textual relay. ViF addresses this by identifying a critical subset of vision tokens with unimodal attention peaks in middle transformer layers and introducing a Visual Flow mechanism to propagate these tokens directly across agent turns, supplemented by an attention reallocation mechanism. The method is model-agnostic and consistently improves performance across eight benchmarks, four MAS topologies, and ten base VLMs.

Problem

Existing VLM-based MAS pipelines relay inter-agent visual information exclusively through generated text (textual flow). This visual-to-text compression causes two compounding failures: (1) intrinsic per-agent hallucination and (2) hallucination propagation, where subsequent agents treat prior hallucinated text as authoritative visual evidence. The result is cascading, self-reinforcing hallucination. Single-agent hallucination-reduction methods cannot address the propagation component, and no prior work formalizes or mitigates this multi-agent-specific failure mode.

Method

The paper first performs turn-, layer-, and token-wise attention analyses on LLaVA-NeXT-7B, revealing that (a) average attention to vision tokens decays from 0.165 at turn 1 to 0.063 at turn 20 (−62%), with the sharpest drop in middle layers (−60%), and (b) a small subset of vision tokens exhibiting a unimodal attention peak in middle layers are disproportionately responsible for visual understanding—their ablation causes ~3× more degradation than dropping equivalent numbers of other token types.

ViF introduces two components:

  1. Visual Information Relay. At each agent handoff, unimodal tokens \(R = \{r_1,\ldots,r_n\} \subset V\) (\(n \ll m\)) are selected by their attention morphology. They are contextualized with instruction tokens via a lightweight transformer block: $\(\hat{R} = f(R \oplus I)[:n]\)$ These relay tokens are inserted between vision and instruction tokens for the next agent, preserving spatial positional encodings.

  2. Attention Reallocation. In middle layers, a temperature-scaled softmax sharpens unimodal patterns: $\(A = \mathrm{Softmax}_\tau(S) = \frac{\exp(s/\tau)}{\sum_i \exp(s_i/\tau)}\)$ Additionally, attention mass is collected from inactive vision tokens and instruction tokens and redistributed to active vision tokens via binary mask matrices \(M_c\) and \(M_r\) (Eqs. 3–4). In deep layers, the reallocation direction reverses (vision → instruction), matching the natural attention flow there.

For models using Flash-Attention 2/3 (where attention scores are inaccessible), a Key-Norm (L2 norm of the key matrix) alternative substitutes for attention-score-based token selection.

Key Contributions

  • Formalization of multi-agent visual hallucination snowballing as a distinct two-mechanism failure (intrinsic hallucination + propagation), separate from single-agent hallucination.
  • Empirical identification of unimodal vision tokens in middle transformer layers as the critical carrier of visual semantics, whose proportion collapses from 1.22% at turn 1 to 0.10% at turn 20.
  • ViF: a lightweight, model-agnostic Visual Flow relay method combining selected unimodal token propagation and layer-specific attention reallocation.
  • A proposed Hallucination Snowballing (HS) metric to quantify multi-agent-specific hallucination accumulation.
  • Comprehensive evaluation: 8 main benchmarks + 4 augmented-visual benchmarks, 4 MAS topologies (linear, layered, random, circular), 10 base VLMs (7B–34B scale), with public code.

Results

  • Average improvement across 8 comprehensive+hallucination benchmarks: +2.4% to +3.8% across all four MAS structures and six 7B base models.
  • On the circular structure with LLaVA-NeXT-34B: +4.4% average improvement; MME gains +25.9, POPE +2.2, CHAIR −2.2, HallBench +2.8.
  • On Qwen2.5-VL-32B (circular): +4.1% average; MME +20.1, POPE +1.5, CHAIR −2.1.
  • HS metric (circular): CHAIR hallucination snowballing reduced from 18.9→12.8 (−32.3%), POPE from 29.1→ (not shown fully), HallBench from 47.4→(not shown); linear structure shows −35.8% average HS reduction, layered −33.6%.
  • Augmented visual benchmarks (multi-image + video, circular): +2.0% average across MMIU, MuirBench, MVBench, Video-MME on four base models.
  • Consistent gains across all structures and models; improvements are additive to already-strong baselines (Qwen2.5-VL-7B, Qwen2-VL-7B).

Limitations

  • The unimodal token selection and attention reallocation rely on access to intermediate attention scores; flash-attention workaround (Key-Norm) is an approximation whose fidelity gap is not fully characterized.
  • Analysis is primarily conducted on LLaVA-NeXT-7B/POPE in the main paper; claims of generalization rest on appendix results.
  • The lightweight transformer block \(f(\cdot)\) adds inference overhead; parameter counts and latency costs are not reported in the main text.
  • Method assumes that visual relay tokens selected at one agent turn remain semantically relevant for the next; how this holds for tasks requiring significant visual re-interpretation is not analyzed.
  • The HS metric is proposed but not yet validated against human judgments of snowballing severity.

Relevance to Agentic AI / LLM Agents

ViF directly addresses a fundamental reliability gap in multi-agent VLM pipelines: the inability of purely text-mediated communication to preserve grounded visual evidence across agent turns, a problem that grows worse with agent depth. For researchers building agentic VLM systems (e.g., visual reasoning chains, multi-agent inspection or captioning pipelines), this provides a model-agnostic plug-in that requires no retraining and integrates into any MAS topology. The attention-level analysis connecting hallucination propagation to systematic visual attention decay offers a mechanistic lens that generalizes beyond VLMs to any multimodal agentic system where inter-agent context compression degrades original modality signals. The work also establishes a new evaluation axis—hallucination snowballing rate—that the agentic AI community should adopt for assessing robustness of multi-turn, multi-agent multimodal workflows.