#plain-text-accounting
Research grounded in plain-text accounting formats and workflows
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.
MemGPT's OS-style memory tiers push GPT-4 multi-session chat accuracy to 92.5% versus 32.1% fixed-context—needed for multi-year ledger agents.
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 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 (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.
Tree of Thoughts (ToT) achieves 74% on Game of 24 vs 4% for standard GPT-4 CoT by organizing LLM reasoning into a branching search tree with pruning and backtracking — with direct implications for multi-step financial classification and tax optimization in Beancount workflows.
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.
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.
ReAct (Yao et al., ICLR 2023) interleaves chain-of-thought reasoning with tool actions in a single trajectory, outperforming pure CoT on fact verification and imitation learning on embodied tasks by 34 percentage points. This analysis covers the paper's failure modes — search-induced distraction and compounding errors — and what they mean for autonomous agents writing back to Beancount ledgers.
A close reading of Toolformer (Meta AI, NeurIPS 2023): how perplexity-filtered self-supervised training teaches a 6.7B-parameter model to call external APIs, where it outperforms GPT-3 175B on arithmetic benchmarks, and why its single-step architecture cannot support the chained tool calls required for structured ledger operations.
