Sssunset/BaRe-Mem-Data

BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation

Overview

In multi-agent systems, a central model can consult advisors, but advisor capabilities vary across tasks, and misleading information can make consultation worse than autonomous reasoning. BaRe-Mem is an online Bayesian reliability memory for multi-agent consultation: it estimates each advisor's reliability from the central model's internal belief representations, updates these estimates from verified outcomes, and uses them to modulate the influence of advisor responses and to choose between consultation and autonomous reasoning.

This repository holds the question streams of the paper. Every question comes with the answers of six advisor models and their verified correctness, in two capability regimes and at five misleading information ratios.

The BaRe-Mem reliability memory update (Figure 1 of the paper). (a) Without verified evidence, all advisors are assigned equal reliability. (b) Updating only candidate 1 with two correct outcomes shifts its estimated reliability to the red posterior, while a subsequent incorrect outcome yields the blue posterior. The annotated gains denote the corresponding Kalman gains.

Dataset at a glance

(a) The capability-supported regime covers mathematical reasoning, retrieval-based question answering and code generation, where most central models are strong and useful advisor information is broadly available. The capability-challenging regime covers physical commonsense, broad knowledge, science questions, complex reasoning and language understanding, where capabilities are far more heterogeneous across models. Within each regime the questions of the benchmarks are shuffled into one stream. (b) As the misleading information ratio grows, advisor answers are replaced by misleading ones: fluent, on-topic and well-formed, but verified to be incorrect. Solid: share of advisor answers that are right; dashed: share of questions with at least one right advisor.

Subset	Questions	Content
capability_supported	4,319	GSM8K test, SQuAD dev and APPS test, with the honest answers of the six advisors
capability_challenging	17,403	PIQA, MMLU, OpenBookQA, SciQ, BBH and SuperGLUE, with the honest answers of the six advisors
capability_supported_misleading_p000 … _p100	4,319 each	The same questions with 0, 25, 50, 75 or 100% of the advisor answers misleading
capability_challenging_misleading_p000 … _p100	17,403 each	The same questions with 0, 25, 50, 75 or 100% of the advisor answers misleading
capability_supported_misleading	25,914	Answer pool: every advisor's verified misleading answer to every question
capability_challenging_misleading	104,418	Answer pool: every advisor's verified misleading answer to every question

Every subset has a single test split. At 100%, a few advisor answers stay right (9.0% of the answers in the capability-supported regime, 3.0% in the capability-challenging regime), where no verified misleading answer was obtained.

Models

The six advisors' answers are part of the data; the six central models read them at evaluation time.

#	Advisor	Central model
1	Gemma-3-4B	Qwen3-4B
2	Phi-4-mini	Qwen3-8B
3	Qwen2.5-Coder-7B	Qwen3-14B
4	Llama-3.1-8B	Ministral-8B
5	DeepSeek-Coder-V2-Lite	Qwen2.5-7B
6	R1-Distill-Qwen-7B	Phi-4

Advisor k is peer_{k-1} in the data (peer_0 … peer_5).

Main result

Accuracy under increasing misleading-advice ratios for Qwen3-14B and Phi-4 (Figure 2 of the paper). Left: capability-supported regime; right: capability-challenging regime. Across both regimes, BaRe-Mem remains above the no-consultation baseline as misleading information increases, while other consultation methods degrade substantially in the capability-challenging regime.

Usage

from datasets import load_dataset

ds = load_dataset("Sssunset/BaRe-Mem-Data", "capability_challenging_misleading_p050",
                  split="test")
q = ds[0]
print(q["problem"], q["answer"])
print(q["peer_responses"]["peer_0"])            # advisor 1's answer
print(q["peer_correct"]["peer_0"])              # its grade (right at >= 0.5)
print(q["peer_metadata"]["peer_0"]["misled"])   # replaced by a misleading answer?

To reproduce the paper, the code downloads these files and checks each against its sha256: python datasets/download.py.

Data fields

Field	Description
id, source	Question id (<benchmark>:<index>) and benchmark
task_type	How answers are graded: math (GSM8K), rag (SQuAD), code (APPS), mcqa, boolqa and shortqa (BBH)
problem, answer	The question and its gold answer
choices, choice_labels	The options of multiple-choice questions
context, answer_aliases	SQuAD's passage and accepted answers
test_cases, code_format, difficulty	APPS's tests, answer format and difficulty
subject, bbh_task	MMLU subject and BBH task
peer_responses	The six advisors' answers, peer_0 … peer_5
peer_correct	Each answer's grade in [0, 1]; an answer counts as right at ≥ 0.5 (correctness_by_peer in the capability-supported subsets: the same as 0/1)
peer_metadata	Each advisor's model and generation settings; in the misleading subsets also regime and misled (the answer was replaced by a verified misleading one)

In the answer pools, each row is one advisor's misleading answer to one question: peer, id, response, its grade (target, correct), whether it passed verification (accepted) and how it was generated (attempts, forced, reasons, soft).

Source benchmarks

The questions come from GSM8K, SQuAD, APPS, PIQA, MMLU, OpenBookQA, SciQ, BBH and SuperGLUE, which keep their original licenses.

Citation

@misc{feng2026baremembayesianreliabilitymemory,
      title={BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation},
      author={Peilin Feng and Zhengyang Huang and Soujanya Poria},
      year={2026},
      eprint={2609.35551},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.35551},
}