fix

Author: codelion

optillm/mars/agent.py

@@ -34,11 +34,11 @@ def _assign_temperature(self) -> float:
     def _get_reasoning_effort(self) -> str:
         """Get reasoning effort level based on agent temperature"""
         if self.temperature <= 0.4:
-            return "low"  # 8k reasoning tokens
+            return "low"  # ~20% of max_tokens for reasoning
         elif self.temperature <= 0.8:
-            return "medium"  # 16k reasoning tokens
+            return "medium"  # ~50% of max_tokens for reasoning
         else:
-            return "high"  # 24k reasoning tokens
+            return "high"  # ~80% of max_tokens for reasoning
 
     def generate_solution(self, problem: str, request_id: str = None) -> Tuple[AgentSolution, int]:
         """Generate a solution for the given problem using reasoning API"""
@@ -51,20 +51,11 @@ def generate_solution(self, problem: str, request_id: str = None) -> Tuple[Agent
             problem=problem
         )
 
-        # Configure reasoning parameters based on fixed budgets
+        # Configure reasoning parameters - simplified with effort only
         reasoning_effort = self._get_reasoning_effort()
-        max_tokens = self.config['max_tokens']  # Fixed 32k
-
-        # Use fixed reasoning tokens based on effort level
-        if reasoning_effort == "low":
-            reasoning_tokens = self.config['low_effort_tokens']  # 8k
-        elif reasoning_effort == "medium":
-            reasoning_tokens = self.config['medium_effort_tokens']  # 16k
-        else:  # high
-            reasoning_tokens = self.config['high_effort_tokens']  # 24k
+        max_tokens = self.config['max_tokens']  # Fixed 30k
 
         reasoning_config = {
-            "max_tokens": reasoning_tokens,
             "effort": reasoning_effort
         }
 
@@ -134,9 +125,8 @@ def verify_solution(self, problem: str, solution: str, verifier_id: int, solutio
             solution=solution
         )
 
-        # Use fixed verification token budgets
-        max_tokens = self.config['max_tokens']  # Fixed 32k
-        verification_reasoning_tokens = self.config['verification_tokens']  # Fixed 8k
+        # Use simplified verification with effort parameter
+        max_tokens = self.config['max_tokens']  # Fixed 30k
 
         try:
             response = self.client.chat.completions.create(
@@ -150,8 +140,7 @@ def verify_solution(self, problem: str, solution: str, verifier_id: int, solutio
                 timeout=180,
                 extra_body={
                     "reasoning": {
-                        "max_tokens": verification_reasoning_tokens,
-                        "effort": "low"
+                        "effort": "low"  # Low effort for verification consistency
                     }
                 }
             )
@@ -196,9 +185,8 @@ def improve_solution(self, problem: str, current_solution: str, feedback: str, i
             issues="\n".join(f"- {issue}" for issue in issues)
         )
 
-        # Use fixed improvement token budgets (use high effort for iterations)
-        max_tokens = self.config['max_tokens']  # Fixed 32k
-        improvement_reasoning_tokens = self.config['high_effort_tokens']  # Fixed 24k
+        # Use simplified improvement with high effort
+        max_tokens = self.config['max_tokens']  # Fixed 30k
 
         try:
             response = self.client.chat.completions.create(
@@ -212,8 +200,7 @@ def improve_solution(self, problem: str, current_solution: str, feedback: str, i
                 timeout=300,
                 extra_body={
                     "reasoning": {
-                        "max_tokens": improvement_reasoning_tokens,
-                        "effort": "high"
+                        "effort": "high"  # High effort for improvements
                     }
                 }
             )

optillm/mars/mars.py

@@ -15,23 +15,17 @@
 
 logger = logging.getLogger(__name__)
 
-# Default MARS configuration with fixed 32k token budget
+# Default MARS configuration - simplified with OpenRouter effort parameter
 DEFAULT_CONFIG = {
     'num_agents': 3,
     'max_iterations': 5,  # Balanced for quality vs efficiency
     'verification_passes_required': 2,  # Balanced for 5-iteration efficiency
     'consensus_threshold': 2,  # Keep at 2 for 3-agent setup
     'min_verified_solutions': 1,  # Keep minimal requirement
-    'max_tokens': 32000,  # Fixed 32k token budget for all calls
+    'max_tokens': 30000,  # Fixed 30k token budget for all calls
     'max_verification_attempts': 3,
     'early_termination': True,
-    'use_reasoning_api': True,
-    # Fixed reasoning token allocations
-    'low_effort_tokens': 8000,     # Agent 0 (temperature 0.3)
-    'medium_effort_tokens': 16000, # Agent 1 (temperature 0.6)
-    'high_effort_tokens': 24000,   # Agent 2 (temperature 1.0)
-    'verification_tokens': 8000,   # Fixed low effort for verification consistency
-    'synthesis_tokens': 24000      # Fixed high effort for final synthesis
+    'use_reasoning_api': True
 }
 
 def multi_agent_reasoning_system(
@@ -189,24 +183,19 @@ def _synthesize_final_solution(
     )
 
     try:
-        # Use fixed synthesis token budgets
-        synthesis_max_tokens = config['max_tokens']  # Fixed 32k
-        synthesis_reasoning_tokens = config['synthesis_tokens']  # Fixed 24k
-
-        # Use fixed reasoning effort for synthesis
+        # Use simplified synthesis with effort parameter
         response = client.chat.completions.create(
             model=model,
             messages=[
                 {"role": "system", "content": "You are a mathematical synthesis expert."},
                 {"role": "user", "content": synthesis_prompt}
             ],
-            max_tokens=synthesis_max_tokens,
+            max_tokens=config['max_tokens'],
             temperature=0.3,  # Lower temperature for synthesis
             timeout=300,
             extra_body={
                 "reasoning": {
-                    "max_tokens": synthesis_reasoning_tokens,
-                    "effort": "high"
+                    "effort": "high"  # High effort for final synthesis
                 }
             }
         )
@@ -219,11 +208,10 @@ def _synthesize_final_solution(
                     {"role": "system", "content": "You are a mathematical synthesis expert."},
                     {"role": "user", "content": synthesis_prompt}
                 ],
-                "max_tokens": synthesis_max_tokens,
+                "max_tokens": config['max_tokens'],
                 "temperature": 0.3,
                 "extra_body": {
                     "reasoning": {
-                        "max_tokens": synthesis_reasoning_tokens,
                         "effort": "high"
                     }
                 }