hallucination, toxic content, faulty code
A problem-solving process is a sequence of iterative stages within a human group (Bransford & Stein, 1993).
this research aims to solve multi-step reasoning taskusing multi-agent framework
Despite the rapid advancement of LLM, its shortcomings of LLMs in areas such as mathematical computations, reasoning, and factuality interrupt application lacks in especially multi-step reasoning task
but existing multi-agent framework
This is due to the overly long token sequences and the interleaved text-image data format, which limit performance.
and their cooperative behaviours with other AI are still not explored well.
prior research has focused on (IOA, evoagent,
multi-agent framework where strong agents supervise weak models demonstrate
Large language models (LLMs) are expected to respond accurately but often exhibit deficient reasoning or generate
hallucinatory content
On the other hand, solving complex reasoning problems requires more in-depth, deliberative and logical thinking steps, i.e., the “System 2" mode [15]. Taking solving math word problems as an example, any incorrect intermediate reasoning step (e.g., calculation errors, mis-interpretations) can potentially lead to incorrect final answers
In multi-agent frameworks (Zeng et al., 2022; Zhuge et al., 2023), several LLMs take on different roles (Li et al., 2023; Park et al., 2023; Qian et al., 2023; Wu et al., 2023), to communicate in natural language and collectively solve a given task. This approach often outperforms single agents, exploiting the specialization (Hong et al., 2023) of various LLM agents. Unfortunately, it also leads to increasingly different and disparate code bases that require a lot of human engineering to define prompting schemes and the workflow of agents.
From the perspective of multi-agent system robustness, any unexpected error or response can propagate to the whole system, causing a series of cascading effects if not handled properly
In the process of completing complex tasks, interactive reasoning among agents within a static MACNET requires strategical traversal to establish an orderly interaction sequence
Soft 프롬프트
Hard prompt is highly sensitive to word choice, where even minor changes can significantly impact generation quality. To address these limitations, soft prompt methods use continuous, trainable vector embeddings, offering more flexible and fine-grained control without altering the underlying model parameters.
task decomposition
self feedback
computation graph optimization
discussion
fall short
suboptimal
challenges
streamline
outline
underline
underscore
facilitate
alleviate
devised
denote, indicate (e.g denoted as, represented by
demonstrate
remarkable
non-trivial
capability
prior
tailored
leverage
prescribe
employ
address
exhibit
desired result
assess
empower
is designed to
elicit
decompose
scalable
reveal that
as shown in
Given
serve as
reliable
uncover
bypass
extensive
emperical
propose
as illustrated in
primarily
substantial
seamlessly integrate
the potential
application in
baseline
refers to (mean)
enable
ensure
compatible with
as follows: (1)
signify
assign
induce
comprehensive
workflow
conduct
tackle
calibrated
discrepancy
investigate
charactarized by
implementation
modular
robust
concurrent
varying (e.g varying levels of expertise)
travelplanning 데이터셋
Quality 데이터셋
power of law와 structure에 관해 MACNET