AI Paper Atlas

Research map / Large language models / Agents, code and math

From Document Layout to Causal Topology: An End-to-End Architecture for Temporal Causal Reasoning

Zhifeng Yang, Wei Luo, Yongjian Chen, Shiqi Ou

NeurIPS 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 0 citations (OpenAlex)

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 reframing temporal causal reasoning from discrete narratives to causal topologies, with a training-free neuro-symbolic agent (TCR-Agent) and a second architecture, based only on the truncated abstract.

Problem

LLMs generally lack deep causal reasoning in temporal reasoning tasks.

Approach

Inspired by document layout reordering, shifts from discrete event narratives to causal topologies; a training-free neuro-symbolic agent (TCR-Agent) performs interpretable causal deduction and counterfactual truncation.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add DAG instantiation to TCR-Agent (ablation)TCR-Agent (GPT-5.4) · TempoCausal-BenchLLM-Acc: 63.1 -> 70.4Improved
add dual-track reasoning to TCR-Agent (ablation)TCR-Agent (GPT-5.4) · TempoCausal-BenchLLM-Acc: 65.2 -> 70.4Improved
add counterfactual intervention to TCR-Agent (ablation)TCR-Agent (GPT-5.4) · TempoCausal-BenchLLM-Acc: 67.5 -> 70.4Improved
add scratchpad to TCR-Agent (ablation)TCR-Agent (GPT-5.4) · TempoCausal-BenchLLM-Acc: 66.8 -> 70.4Improved
add causal-topological attention mask to TCR-Former (ablation)TCR-Former (Qwen-3-8B) · TempoCausal-BenchLLM-Acc: 67.0 -> 75.7Improved
add fact-logic bypass fusion to TCR-Former decoder (ablation)TCR-Former (Qwen-3-8B) · TempoCausal-BenchLLM-Acc: 71.4 -> 75.7Improved
add time span bias to TCR-Former (ablation)TCR-Former (Qwen-3-8B) · TempoCausal-BenchLLM-Acc: 72.3 -> 75.7Improved
add a causal-topological encoder with a Qwen-3-0.6B backbone without the proposed mask and biases (Only Encoder + Decoder) versus decoder alone (ablation)TCR-Former (Qwen-3-8B) · TempoCausal-BenchLLM-Acc: 58.6 -> 55.4Hurt
replace external prompt methods (TG-LLM, Time-CoT, TISER) with TCR-AgentGPT-5.4 · TempoCausal-BenchLLM-Acc: 59.2 -> 70.4Improved
internalize causal topology into the model (TCR-Former) rather than LoRA fine-tuning on TCR-Agent outputsQwen-3-8B · TempoCausal-BenchLLM-Acc: 67.4 -> 75.7Improved
add TCR-Agent external graph and CoT reasoning (versus TISER) for rule inference errorsQwen-3-8B · TempoCausal-Bench error attributionRule inference mismatch error proportion: 13.8 (TISER) -> 13.0 (TCR-Agent)No clear effect
use TCR-Former instead of TCR-Agent for inference costQwen-3-8B · TempoCausal-Benchinference latency: -62% latency, about -44% output tokensImproved

Limitations

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy