Introduction
When my AI course professor assigned us to read and blog about recent research
papers, I expected to spend a weekend forcing myself through dense academic
writing. What actually happened shifted how I think about two topics I thought
I already understood: search algorithms and intelligent agents.
The two papers I chose:
- "A Survey of LLM-based Deep Search Agents" (2026)
- "Research on the A* Algorithm Based on Adaptive Weights" (2025)
are not obviously related at first glance. One is about large language models
navigating information, the other is about a classic pathfinding algorithm.
But together, they tell a unified story about what search really means
in modern AI.
Paper 1: A Survey of LLM-based Deep Search Agents (2026)
What is the paper trying to do?
This survey maps out an entirely new category of AI system: the Deep Search
Agent. The authors document how we evolved from:
Traditional keyword search → LLM-enhanced search → Agentic search
In agentic search, an LLM autonomously plans, retrieves, reflects, and
iterates until it can synthesize a comprehensive answer to a complex question.
The core architectural loop the paper formalizes is:
Plan → Act → Observe → Reflect → Generate
Crucially, the Reflect stage decides whether to loop back and search
again or to finalize the answer. Real-world systems like OpenAI's Deep
Research and Perplexity's Pro Search are concrete implementations
of this loop.
The Technical Depth
These agents organize their retrieval using three distinct search strategies:
| Strategy | How it works |
|---|---|
| Sequential | Works through one reasoning thread at a time, step by step |
| Parallel | Decomposes the question into sub-questions, explores all branches simultaneously |
| Hybrid | Combines both, sometimes using Monte Carlo Tree Search (MCTS) |
The task formulation is a state-space problem: given a user query q,
the agent operates over a trajectory of (observation, action) pairs and must
reach a terminal state where the accumulated evidence satisfies the
information need.
Connection to Our Course
In our AI course, we study rational agents using the PEAS framework
(Performance, Environment, Actuators, Sensors). A Deep Search Agent is
structurally a goal-based agent where:
- Goal — answering the query completely
- Actuator — ability to query search engines or browse URLs
- Sensor — the retrieved text it reads
- Environment — the open web (partially observable, dynamic)
The search strategies map directly to what we study in class:
Sequential search = Depth-First Search (DFS)
Parallel search = Breadth-First Search (BFS)
The goal test = the moment the agent decides it has enough information
Paper 2: A* Algorithm Based on Adaptive Weights (2025)
What is the paper trying to do?
The standard A* algorithm has a well-known limitation: it expands a large
number of nodes trying to be both fast and optimal simultaneously.
This paper proposes a solution — a dynamic weighting function based on
a radial basis function — to make A* smarter about when to rush and when
to be careful.
The core formula becomes:
f(n) = g(n) + f(x,y) × h(n)
Where f(x,y) is not a constant but a dynamic weight that changes based
on the agent's position:
| Position | Weight | Behaviour |
|---|---|---|
| Near start | High | Heuristic amplified, algorithm charges quickly toward goal |
| Near goal | Low | Algorithm slows down, searches carefully for optimal path |
Secondary Optimization
The paper also adds two post-processing steps:
- Removing redundant intermediate nodes — eliminates unnecessary waypoints
- Bessel curve smoothing — makes the final path practical for real robots and vehicles
Results
Compared to standard A*, the adaptive version reduces nodes searched by an
average of 76.4% and reduces total turn angle by 71.7%. For autonomous
vehicles and robots, this means less computation and smoother, safer paths.
Connection to Our Course
In our lectures on informed search, we study A* as the gold standard —
guaranteed to find the optimal solution if h(n) is admissible (never
overestimates the true cost).
This paper asks a deeper question:
Even within admissible heuristics, can we be smarter about how much we
trust the heuristic at different points in the search?
The answer is yes — and it maps directly to our course concept of
weighted A. What makes this paper genuinely interesting is that the
weight is not a fixed constant chosen by a human engineer, but a function
of the agent's position, computed automatically. The **Euclidean distance
heuristic* we study in class is retained, but modulated dynamically rather
than applied uniformly.
My NotebookLM Experience vs. Manual Reading
I read both papers manually first, and it was harder than expected. The
taxonomy tables in the survey paper required multiple passes. The radial
basis function notation in the A* paper was not immediately clear. But
that struggle built real understanding.
When I then used NotebookLM, it was useful for generating follow-up
questions and locating specific sections quickly — but it simplified
concepts I now knew were more nuanced. It flattened the dynamic weighting
idea into something almost trivial compared to what the paper actually
presents.
My recommendation: read the paper yourself first, then use NotebookLM
to fill gaps and test your understanding — not to replace the reading.
Conclusion
I came into this assignment thinking of search algorithms and agents as
separate textbook chapters. I am leaving it understanding them as two
perspectives on the same underlying problem: how does an intelligent system
efficiently navigate a space of possibilities to reach a goal?
Whether that space is a city map, the open web, or a robot's environment,
the core problem is identical. The adaptive A* paper makes the heuristic
context-sensitive. The Deep Search Agent paper replaces a hand-coded
heuristic with an LLM's judgment about relevance. Both are moving in the
same direction — search that adapts to where it is, rather than applying
the same strategy everywhere blindly.
cc: @raqeeb_26
Paper references:
Xi et al. (2026). A Survey of LLM-based Deep Search Agents. arXiv:2508.05668
Liu et al. (2025). Research on the A Algorithm Based on Adaptive Weights.*
Top comments (0)