ReflAct - Reflection-Grounded Agent Reasoning¶

Theory¶

Paper¶

Original paper

Kim, J., Rhee, S., Kim, M., et al. (2025) ReflAct: World-Grounded Decision Making in LLM Agents via Goal-State Reflection DOI: 10.48550/arXiv.2505.15182 EMNLP 2025 (Main Conference)

Concept

ReflAct extends ReAct with state‑grounded reflection. Instead of forward‑planning ("What should I do next?"), the agent reflects on its current state relative to the goal ("Where am I relative to the goal?"). This enables systematic self‑correction and better goal focus.

Architecture¶

flowchart LR
    query([Query]) --> service[AgentChatService]
    service --> reflact[ReflAct Mode]
    reflact --> refl[REFLECTION]
    refl --> action[ACTION]
    action --> tool[[Tool Execute]]
    tool --> obs[/OBSERVATION/]
    obs --> decide{Goal reached?}
    decide -->|No| refl
    decide -->|Yes| final([FINAL ANSWER])

    style query fill:#98d4bb,stroke:#6bbf9a,color:#000
    style service fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style reflact fill:#88c4c8,stroke:#5fa8ad,color:#000
    style refl fill:#88c4c8,stroke:#5fa8ad,color:#000
    style action fill:#D1BC8A,stroke:#b8a06a,color:#000
    style tool fill:#b0ca97,stroke:#8fb077,color:#000
    style obs fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style decide fill:#9e9e9e,stroke:#757575,color:#fff
    style final fill:#e8c87a,stroke:#d4a84b,color:#000

ReflAct loop: Query → REFLECTION (state‑grounded) → ACTION → Tool → OBSERVATION → (repeat or answer)

Core Concept¶

REFLECTION → ACTION → OBSERVATION → REFLECTION → ... → FINAL ANSWER

The REFLECTION step is state‑grounded:

"Where am I relative to the goal?"
"What do I already know?"
"What did I just discover?"
"What is still missing to reach the goal?"

Difference from ReACT¶

Aspect	ReACT (THOUGHT)	ReflAct (REFLECTION)
Focus	Forward‑looking	State‑grounded
Question	"What should I do next?"	"Where am I relative to the goal?"
Perspective	Planning	Evaluation
State awareness	Partial	Full
Self‑correction	Possible	Systematic

REFLECTION Structure¶

Each reflection contains three components:

REFLECTION:
1. Current state: [What I already know]
2. Latest discovery: [What was just found]
3. Goal relation: [How close am I? What is still missing?]

Advantages over ReACT¶

Advantage	Description
Better self‑correction	Explicit state evaluation enables early error detection
Goal focus	Each step is measured against the goal
Transparency	Fully traceable reasoning process
Robustness	Less prone to drifting off goal

Example¶

Query¶

"When is DG Agentur open and how can I reach them?"

ReflAct Trace¶

[Iteration 1]
REFLECTION:
  Current state: The user asks for opening hours AND contact details.
                 I have no information yet.
  Latest discovery: (None - start of task)
  Goal relation: I need both pieces of information. I will search
                 for opening hours first.
ACTION: rag_search("opening hours DG Agentur")
OBSERVATION:
  [1] Contact: Mon-Fri 9-18, Sat by appointment
  [2] You can reach us at +49 911 97554990

[Iteration 2]
REFLECTION:
  Current state: I have opening hours (Mon-Fri 9-18, Sat by appointment)
                 and a phone number (+49 911 97554990).
  Latest discovery: Document [1] contains opening hours,
                    [2] provides the phone number.
  Goal relation: Opening hours: ✓ complete
                 Contact details: partial (phone ✓, email missing, address missing)
ACTION: rag_search("contact email address legal notice")
OBSERVATION:
  [1] Legal notice: Phone +49 911 97554990, Email office@dg-agentur.de
  [2] Address: Bahnhofstraße 2, 90402 Nürnberg

[Iteration 3]
REFLECTION:
  Current state: I have all required information:
                 - Opening hours: Mon-Fri 9-18, Sat by appointment
                 - Phone: +49 911 97554990
                 - Email: office@dg-agentur.de
                 - Address: Bahnhofstraße 2, 90402 Nürnberg
  Latest discovery: Email and address from the legal notice.
  Goal relation: GOAL REACHED - all information complete.
FINAL ANSWER: DG Agentur is open Monday to Friday 9-18,
              Saturday by appointment. [1]

              Contact:
              - Phone: +49 911 97554990
              - Email: office@dg-agentur.de
              - Address: Bahnhofstraße 2, 90402 Nürnberg [2]

Response¶

DG Agentur is open Monday to Friday 9-18,
Saturday by appointment. [1]

Contact:
- Phone: +49 911 97554990
- Email: office@dg-agentur.de
- Address: Bahnhofstraße 2, 90402 Nürnberg [2]

Sources:
[1] Contact - DG Agentur
[2] Legal notice - DG Agentur

Implementation in LLARS¶

Status: Production

ReflAct is fully implemented and in production use.

Architecture¶

flowchart TB
    subgraph main[chat_reflact Loop]
        direction TB
        start([Start]) --> refl[Generate REFLECTION]
        refl --> action[Generate ACTION]
        action --> parse[Parse Response]
        parse --> check{FINAL ANSWER?}
        check -->|No| exec[Execute Tool]
        exec --> history[Add to Steps]
        history --> refl
    end
    check -->|Yes| done([Finalize])

    style start fill:#98d4bb,stroke:#6bbf9a
    style refl fill:#88c4c8,stroke:#5fa8ad,color:#000
    style action fill:#D1BC8A,stroke:#b8a06a,color:#000
    style parse fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style check fill:#9e9e9e,stroke:#757575,color:#fff
    style exec fill:#b0ca97,stroke:#8fb077,color:#000
    style history fill:#88c4c8,stroke:#5fa8ad,color:#000
    style done fill:#e8c87a,stroke:#d4a84b
    style main fill:#f5f5f5,stroke:#88c4c8

System Prompt¶

# DEFAULT_REFLACT_SYSTEM_PROMPT (db/models/chatbot.py)
"""
Du bist ein ReflAct-Agent. Bei jedem Schritt reflektierst du deinen aktuellen Zustand
RELATIV zum Aufgabenziel, dann wählst du die nächste Aktion.

## ReflAct-Prinzip (basierend auf arxiv.org/abs/2505.15182):
- Nicht "Was soll ich als nächstes tun?" (vorausschauend)
- Sondern "Wo stehe ich relativ zum Ziel?" (zustandsbasiert)

## Deine Reflection muss IMMER enthalten:
1. Aktueller Zustand: Was weißt du bereits?
2. Letzte Entdeckung: Was hast du gerade erfahren?
3. Ziel-Relation: Wie nah bist du dem Ziel? Was fehlt noch?

## Verfügbare Aktionen:
- rag_search("suchbegriff") - Semantische Dokumentensuche
- lexical_search("suchbegriff") - Keyword-Suche

## Format (STRIKT einhalten!):

REFLECTION: Aktuell weiß ich [Zustand]. Die letzte Suche ergab [Ergebnis]. Dies bringt mich [näher/nicht näher] zum Ziel [X], weil [Begründung].
ACTION: rag_search("suchbegriff")

Wenn das Ziel erreicht ist:
REFLECTION: Ich habe alle nötigen Informationen: [Zusammenfassung]. Das Ziel ist erreicht.
FINAL ANSWER: [Vollständige Antwort basierend auf den gefundenen Informationen]
"""

Additionally: - chatbot.system_prompt is prefixed. - build_tool_availability_prompt() adds the enabled tools dynamically. - {PROJECT_URL} placeholders are replaced before use.

Files¶

File	Function
`app/services/chatbot/agent_chat_service.py`	Routing to ACT/ReAct/ReflAct
`app/services/chatbot/agent_modes/mode_reflact.py`	`chat_reflact()` loop + streaming
`app/services/chatbot/agent_parsers.py`	`parse_reflact_response_v2()`
`app/services/chatbot/agent_tools.py`	Tool execution + confidence checks
`app/db/models/chatbot.py`	DEFAULT_REFLACT_SYSTEM_PROMPT + prompt settings

Code Snippet¶

# mode_reflact.py - chat_reflact()
for iteration in range(max_iterations):
    yield {"status": "iteration", "iteration": iteration + 1, "max": max_iterations, "goal": goal}

    # Stream REFLECTION + ACTION
    response_text, reflection, action, final_answer = yield from _stream_reflact_response(...)

    if final_answer:
        yield {"status": "final_answer"}
        ...
        return

    # Execute tool
    result, sources = service._tool_executor.execute_tool(action_name, action_param, message, enabled_tools)
    yield {"status": "observation", "result_preview": result[:300], "iteration": iteration + 1}

Parsing¶

# agent_parsers.py - parse_reflact_response_v2()
REFLECTION_PATTERN = r"REFLECTION:\s*(.+?)(?=ACTION:|FINAL ANSWER:|THOUGHT:|GOAL:|$)"
ACTION_PATTERN = r"ACTION:\s*(.+?)(?=OBSERVATION:|REFLECTION:|FINAL ANSWER:|THOUGHT:|GOAL:|$)"
FINAL_PATTERN = r"FINAL ANSWER:\s*(.+?)(?=ACTION:|REFLECTION:|THOUGHT:|GOAL:|$)"

# Backward‑compatible: THOUGHT is interpreted as REFLECTION
THOUGHT_AS_REFLECTION = r"THOUGHT:\s*(.+?)(?=ACTION:|FINAL ANSWER:|REFLECTION:|GOAL:|$)"

Configuration¶

# ChatbotPromptSettings
agent_mode: str = "reflact"
task_type: str = "lookup" | "multihop"
agent_max_iterations: int = 5

# Multihop: max_iterations = min(agent_max_iterations + 2, 10)

tools_enabled: List[str] = ["rag_search", "lexical_search", "respond"]
web_search_enabled: bool = False
web_search_max_results: int = 5

reflact_system_prompt: str = "..."  # custom prompt (optional)

Adaptive Iteration (High Confidence)¶

If the search yields high confidence, ReflAct exits early and generates a final answer immediately. Confidence is derived from source scores (check_high_confidence).

Events (WebSocket)¶

# Streaming Events (excerpt)
yield {"status": "starting", "mode": "reflact"}
yield {"status": "iteration", "iteration": 1, "max": 7, "goal": "...", "steps": [...]}
yield {"status": "reflecting", "iteration": 1}
yield {"status": "reflection_delta", "delta": "...", "iteration": 1}
yield {"status": "reflection", "reflection": "...", "iteration": 1}
yield {"status": "action_delta", "delta": "...", "iteration": 1}
yield {"status": "action", "action": "rag_search", "param": "...", "iteration": 1}
yield {"status": "observation_delta", "delta": "...", "iteration": 1}
yield {"status": "observation", "result_preview": "...", "iteration": 1}
yield {"status": "adaptive_iteration", "iteration": 1, "reason": "high_confidence"}
yield {"status": "adaptive_response", "reason": "high_confidence_results"}
yield {"status": "max_iterations_reached"}
yield {"status": "final_answer"}
yield {"delta": "..."}
yield {"done": True, "full_response": "...", "sources": [...], "goal": "..."}

Logs¶

[AgentChatService] ReflAct adaptive iteration: high confidence on iteration 2

Comparison: ReACT vs ReflAct in LLARS¶

Aspect	ReACT	ReflAct
Method	`chat_react()`	`chat_reflact()`
Location	`mode_react.py`	`mode_reflact.py`
Reasoning step	THOUGHT (forward)	REFLECTION (state‑grounded)
Parsing	`parse_react_response()`	`parse_reflact_response_v2()`
State tracking	Implicit	Explicit (3‑part structure)
Goal evaluation	No	Yes (goal relation)
Tokens/iteration	~150-300	~200-400
Typical iterations	2-5	2-5
Self‑correction	Possible	Systematic

When to use ReflAct over ReACT¶

Use case	Recommendation
Simple lookups	ReACT or ACT
Multi‑hop with clear goal	ReflAct
Complex research	ReflAct
Maximum transparency	ReflAct
Self‑correction important	ReflAct
Minimal token usage	ACT or ReACT

ReflAct - Reflection-Grounded Agent Reasoning¶

Theory¶

Paper¶

Architecture¶

Core Concept¶

Difference from ReACT¶

REFLECTION Structure¶

Advantages over ReACT¶

Example¶

Query¶

ReflAct Trace¶

Response¶

Implementation in LLARS¶

Architecture¶

System Prompt¶

Files¶

Code Snippet¶

Parsing¶

Configuration¶

Adaptive Iteration (High Confidence)¶

Events (WebSocket)¶

Logs¶

Comparison: ReACT vs ReflAct in LLARS¶

When to use ReflAct over ReACT¶

See Also¶