AgentBench：评估作为代理的 LLM —— 对金融 AI 可靠性的启示
AgentBench（Liu 等人，ICLR 2024）在 8 个交互式环境中对 27 个大语言模型进行了基准测试 —— GPT-4 的综合得分为 4.01，而表现最好的开源模型仅为 0.96。三种主要的失败模式（知识图谱失败中 67.9% 为超出任务限制、数据库失败中 53.3% 为格式错误以及无效操作）直接对应了在真实账本上部署 Beancount 回写代理的风险。
BloombergGPT and the Limits of Domain-Specific LLMs in Finance
Bloomberg trained a 50B-parameter LLM on 569B tokens of financial data and beat general models on sentiment and table-reasoning benchmarks — then GPT-4 matched it without any finance-specific pretraining. What the $10M experiment reveals about domain pretraining trade-offs, tokenization of numbers, and why tool-use is more reliable than model internals for accounting agents.
AutoGen: Multi-Agent Conversation Frameworks for Finance AI
AutoGen (Wu et al., 2023) introduces a multi-agent conversation framework where LLM-backed agents pass messages to complete tasks; a two-agent setup lifts MATH benchmark accuracy from 55% to 69%, and a dedicated SafeGuard agent improves unsafe-code detection by up to 35 F1 points — findings directly applicable to building safe, modular Beancount automation pipelines.
Gorilla: How Retrieval-Aware Training Reduces LLM API Hallucinations from 78% to 11%
Gorilla (Patil et al., NeurIPS 2024) fine-tunes a 7B LLaMA model with Retriever-Aware Training on retrieved API documentation, cutting hallucination rates from 78% to 11% versus GPT-4 zero-shot — with direct implications for finance AI write-back agents where wrong account names or inverted signs are correctness failures, not annoyances.
MemGPT: Virtual Context Management for LLM Agents
MemGPT applies OS-style virtual memory paging to LLMs, using three-tier storage — working memory, recall, and archival — to give agents persistent recall across sessions; on multi-session chat benchmarks, MemGPT with GPT-4 achieves 92.5% accuracy versus a 32.1% fixed-context baseline.
SWE-agent: How Interface Design Unlocks Automated Software Engineering
SWE-agent (NeurIPS 2024) introduces Agent-Computer Interfaces (ACIs) — purpose-built layers between LLMs and software environments — showing a 10.7-percentage-point improvement over raw shell access and 12.47% resolution on SWE-bench with GPT-4 Turbo. Interface design, not model capability, is the primary bottleneck for autonomous coding agents.
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
SWE-bench evaluates language models on 2,294 real GitHub issues across 12 Python repositories using execution-based tests; at publication, Claude 2 resolved only 1.96% of issues with realistic retrieval, establishing the de facto benchmark for coding agents and revealing retrieval and patch-length failure modes directly relevant to Beancount write-back agents.
CodeAct: Why Executable Python Code Makes LLM Agents 20% More Accurate
CodeAct (ICML 2024) replaces JSON tool-calling with executable Python code, improving GPT-4 agent success rates by ~20 percentage points on multi-tool tasks and reducing interaction turns by 30% — with direct implications for building reliable Beancount reconciliation agents.
LLMs Cannot Self-Correct Reasoning Yet — ICLR 2024 Findings and Finance AI Implications
Huang et al. (ICLR 2024) show that LLMs asked to review their own reasoning without external feedback consistently degrade accuracy — GPT-4 drops from 95.5% to 91.5% on GSM8K — and what this means for designing reliable Beancount journal entry agents.
Reflexion: Language Agents That Learn from Mistakes Without Retraining
Reflexion (NeurIPS 2023) lets LLM agents improve by storing verbal post-mortems in an episodic buffer — no weight updates required. It reaches 91% on HumanEval with GPT-4 but fails on WebShop, revealing a structural constraint: verbal reinforcement only works when the evaluator produces a crisp, actionable signal. Here is what that means for building a self-correcting Beancount ledger agent.
PAL: Program-Aided Language Models for Reliable Financial Arithmetic
PAL (Program-Aided Language Models) achieves a +38pp accuracy gain over chain-of-thought on arithmetic-heavy tasks by delegating computation to a Python interpreter — a directly applicable architecture for reliable Beancount ledger queries and finance AI.
Can LLMs Reason Over Tabular Data? What Four Benchmarks Tell Us About Finance AI
Four 2024–2025 benchmarks show GPT-4 scoring 42% on real-world table QA versus 86% for humans, with complex aggregations collapsing to 19.6%—and Beancount's native syntax sits at the worst-performing end of the serialization hierarchy for LLM input.