Research map / Multi-agent systems and games / LLM agents
BRIDGE: Triangular Fixed-Point Refinement for Long-Horizon Persona Consistency
Yinghui Jiang, Bocheng Xu, Jianye Xie, Haotong Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BRIDGE, a triangular fixed-point refinement method to keep utterances, internal representations, and memory consistent across long-horizon persona dialogues.
Problem
Long-horizon dialogue agents suffer latent state drift, where what the agent says, internally represents, and stores in memory diverge across turns, causing persona contradictions.
Approach
Performs triangular fixed-point refinement jointly over the agent's utterances, internal latent state, and memory store so that the three stay mutually consistent, with gated dynamic evolution (per the method name); details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add triangular fixed-point refinement (vs standard attention) (ablation) | BRIDGE · CoSER | avg score: 54.6 -> 59.5 | Improved |
| add hierarchical memory evolution (ablation) | BRIDGE · CoSER | avg score: 55.2 -> 59.5 | Improved |
| add personality memory tier (ablation) | BRIDGE · CoSER | avg score: 53.4 -> 59.5 | Improved |
| add dual-system processing (ablation) | BRIDGE · CoSER | avg score: 57.9 -> 59.5 | Improved |
| use K=3 refinement steps instead of one (ablation) | BRIDGE · CoSER | avg score: 58.3 -> 59.5 | Improved |
| use full O-L-M coupling instead of only O<->L (ablation) | BRIDGE · CoSER | avg score: 56.5 -> 59.5 | Improved |
| use learned fast/slow router instead of linear combination (ablation) | BRIDGE · CoSER | avg score: 56.8 -> 59.5 | Improved |
| use all three memory tiers instead of only affective tier (ablation) | BRIDGE · CoSER | avg score: 50.8 -> 59.5 | Improved |
| add triangular fixed-point refinement (ablation) | BRIDGE · PersonaGym | avg score: 4.49 -> 4.59 | Improved |
| add chain-of-thought prompting to GPT-4o | GPT-4o · CoSER | avg score: 55.4 -> 56.7 | Improved |
| backbone-matched BRIDGE versus LoRA (r=64) | BRIDGE · CoSER | avg score: 55.7 -> 59.5 | Improved |
| add BRIDGE to Qwen2.5-32B-Instruct | BRIDGE · PersonaGym | Toxicity Control | Hurt |
Closest papers by meaning
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)
- Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.87)
- Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory (NeurIPS 2026 · similarity 0.87)
- Beyond Prompts: Dynamic Conversational Benchmarking of Large Language Models (NeurIPS 2024 · similarity 0.86)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
