Marketing Manager
τ²-bench extends agent benchmarking to dual-control settings where both the AI and the user invoke tools over shared state — finding that active users cut success rates by 18–25 percentage points, with direct implications for Beancount agents sharing write access with human users.
WorkArena++ (NeurIPS 2024) benchmarks 682 compositional enterprise tasks across three difficulty levels. GPT-4o solves 2.1% of them while humans solve 93.9%, isolating exactly why current AI agents fail at implicit-goal knowledge work and why that gap matters for autonomous accounting automation.
GAIA benchmarks 466 real-world tasks across three difficulty levels; frontier agents reached 74.55% in mid-2026 versus 92% for humans, and the remaining Level 3 gap maps directly to the multi-step coordination challenges in automated Beancount ledger workflows.
OSWorld (NeurIPS 2024) benchmarks multimodal AI agents on 369 real desktop tasks across Ubuntu, Windows, and macOS — finding a 60-percentage-point gap between the best model (12.24%) and human performance (72.36%), with 75% of failures traced to visuomotor grounding errors rather than reasoning failures.
GPT-4 completes only 14.41% of WebArena's 812 realistic web tasks while humans reach 78.24%; the dominant failure mode is false infeasibility — conservative refusal to act — with direct implications for any agent operating Fava or finance web UIs.
WorkArena benchmarks LLM web agents on 33 real ServiceNow tasks — GPT-4o reaches 42.7% overall but 0% on list-filter tasks, exposing a hard wall between form-filling and structured UI interaction that maps directly to challenges in Beancount ledger automation.
τ-bench shows that top LLMs like Claude 3.5 Sonnet drop from pass@1 of 0.692 to pass@4 of 0.462 in retail customer-service tasks — a consistency cliff with direct implications for any write-back agent operating on a Beancount ledger.
Chain-of-Table (ICLR 2024) improves LLM tabular reasoning by evolving the table itself as the intermediate state — achieving 67.31% on WikiTQ vs. 61.48% for prior baselines, with a +10.25 point advantage on tables exceeding 4,000 tokens and direct applicability to Beancount ledger query agents.
TableLlama fine-tunes Llama 2 (7B) on 2.6M table-task examples and beats GPT-4 on structural tasks like column type annotation (F1 94 vs 32), but falls 33 points short on WikiTQ compositional reasoning — a calibrated benchmark for what 7B open models can and cannot do in finance AI today.
TAPAS (Google Research, ACL 2020) answers table questions by selecting cells and applying scalar aggregations — no SQL generated. This post analyzes the architecture, its 12-point SQA accuracy gain, and why the cell-selection paradigm fits small Beancount ledger queries but breaks down at scale.
MAC-SQL (COLING 2025) uses three specialized agents — Selector for schema reduction, Decomposer for question decomposition, and Refiner for execution-guided SQL correction — to reach 59.59% execution accuracy on the BIRD benchmark; ablation shows the Refiner contributes the most (+4.63 points), with direct implications for Beancount ledger query generation.
DIN-SQL (NeurIPS 2023) decomposes text-to-SQL into schema linking, complexity classification, and SQL generation stages, lifting GPT-4 from 67.4% to 85.3% execution accuracy on Spider without fine-tuning — and the same decomposition strategy maps directly onto natural language interfaces for Beancount's BQL query language.
