# Доверие

Надеждност, калибриране и халюцинации във финансови ИИ системи

8 публикации [Вижте всички етикети](https://beancount.io/bg/bean-labs/research-logs/tags) [Обратно към всички публикации](https://beancount.io/bg/bean-labs/research-logs)

9 юли 2026 г. 6 минути четене

## [LLM Confidence and Calibration: A Survey of What the Research Actually Shows](https://beancount.io/bg/bean-labs/research-logs/2026/07/09/confidence-estimation-calibration-llms-survey)

Verbalized GPT-4 confidence hits only \~62.7% AUROC, barely above chance. Uncertainty-aware finance agents need better calibration than that.

Mike Thrift

llm ai machine-learning

1 юли 2026 г. 6 минути четене

## [Uncertainty-Aware Deferral for LLM Agents: When to Escalate from Small to Large Models](https://beancount.io/bg/bean-labs/research-logs/2026/07/01/redact-uncertainty-aware-deferral-llm-agents)

ReDAct defers from a small model to a large one only when perplexity signals uncertainty, cutting cost 64% at matching accuracy for agent workflows.

Mike Thrift

ai llm automation

5 юни 2026 г. 6 минути четене

## [Verifiably Safe Tool Use for LLM Agents: STPA Meets MCP](https://beancount.io/bg/bean-labs/research-logs/2026/06/05/verifiably-safe-tool-use-llm-agents-stpa-mcp)

STPA plus capability-enhanced MCP yields formal safety specs for LLM tool use, with Alloy proving no unsafe flows in a calendar case study.

Mike Thrift

ai llm security

29 май 2026 г. 6 минути четене

## [AGrail: Adaptive Safety Guardrails for LLM Agents That Learn Across Tasks](https://beancount.io/bg/bean-labs/research-logs/2026/05/29/agrail-lifelong-agent-guardrail-adaptive-safety-detection)

AGrail's two-LLM guardrail cuts prompt injection attack success to 0% while preserving 95.6% of benign agent actions on Safe-OS.

Mike Thrift

ai llm security

28 май 2026 г. 6 минути четене

## [ShieldAgent: Verifiable Safety Policy Reasoning for LLM Agents](https://beancount.io/bg/bean-labs/research-logs/2026/05/28/shieldagent-verifiable-safety-policy-reasoning-llm-agents)

ShieldAgent hits 90.4% accuracy on agent attacks with 64.7% fewer API calls by using probabilistic rule circuits instead of LLM guardrails.

Mike Thrift

ai llm machine-learning

25 май 2026 г. 6 минути четене

## [GuardAgent: Deterministic Safety Enforcement for LLM Agents via Code Execution](https://beancount.io/bg/bean-labs/research-logs/2026/05/25/guardagent-safeguard-llm-agents-guard-agent-knowledge-enabled-reasoning)

GuardAgent enforces LLM agent policies by running Python code, hitting 98.7% accuracy with no task failures, versus 81% and up to 71% failure for prompt rules.

Mike Thrift

ai llm automation

28 април 2026 г. 6 минути четене

## [Self-Correction: GPT-4 falls 95.5%→91.5% on GSM8K](https://beancount.io/bg/bean-labs/research-logs/2026/04/28/llms-cannot-self-correct-reasoning-yet)

Huang et al. (ICLR 2024): intrinsic self-correction drops GPT-4 from 95.5% to 91.5% on GSM8K—ledger agents need external validators, not self-review.

Mike Thrift

llm ai machine-learning

19 април 2026 г. 6 минути четене

## [PHANTOM (NeurIPS 2025): Измерване на откриването на халюцинации в LLM при финансови документи](https://beancount.io/bg/bean-labs/research-logs/2026/04/19/phantom-hallucination-detection-financial-long-context)

PHANTOM (NeurIPS 2025) измерва откриването на халюцинации в реални SEC документи. Qwen3-30B води с F1=0.882, но 7B моделите гадаят почти случайно.

Mike Thrift

llm ai machine-learning

[Започни безплатно→](https://beancount.io/sign-up/?next=%2Fledger%2F) [Документация](https://beancount.io/bg/docs)

## Where to look next

- [Sitemap](https://beancount.io/sitemap.xml) - all pages with hreflang alternates
- [LLMs index](https://beancount.io/llms.txt) - curated links for AI agents (also /llms-full.txt for full docs)
- [Documentation](https://beancount.io/docs) - every doc is also available as text/markdown via .md suffix or Accept negotiation
- [REST API reference](https://beancount.io/docs/api-reference) - interactive endpoints, schemas, authentication, and code samples
- [OpenAPI spec](https://beancount.io/openapi.json) - OpenAPI 3.1 spec for this website's public content endpoints; the product contract for ledger operations is at https://beancount.io/api-gateway/v1/openapi.json
- [GraphQL](https://beancount.io/graphql) - public, read-only content API over the blog and research logs (no credentials, no ledger data): 12 query fields over 23 types
- [MCP server](https://beancount.io/mcp) - Model Context Protocol at https://beancount.io/api-gateway/mcp
