| name | on-uncertainty-model-based-multi-agent |
| description | Apply entropy-based uncertainty analysis to multi-agent LLM systems. Diagnose when multi-agent setups hurt performance, select the best solution from multiple agent runs using the Entropy Judger technique, and design uncertainty-aware agent topologies. Trigger phrases: 'analyze multi-agent uncertainty', 'entropy judger for agent selection', 'should I use multiple agents or one', 'select best answer from agent runs', 'diagnose multi-agent failure', 'optimize agent topology with entropy' |
Entropy-Based Uncertainty Analysis for Multi-Agent LLM Systems
This skill enables Claude to apply the Entropy Judger framework from Zhao et al. (2026) to diagnose, evaluate, and improve multi-agent LLM systems. The core technique uses token-level, trajectory-level, and round-level entropy measurements to predict whether a multi-agent configuration will outperform a single agent, select the best solution from multiple runs (pass@k), and identify failure modes early in agent interactions. A critical finding: a single agent outperforms multi-agent systems in ~43% of cases, and first-round entropy dynamics are the strongest predictor of final correctness.
When to Use
- When the user is building a multi-agent system and wants to know whether multiple agents will actually help versus a single agent
- When the user has multiple candidate outputs from agent runs and needs to select the best one without ground truth
- When debugging why a multi-agent pipeline (debate, sequential, centralized) produces worse answers than expected
- When designing agent topologies (sequential, centralized, decentralized, debate, hybrid) and choosing which fits a task
- When the user wants to add an entropy-based quality gate to filter or rank LLM outputs before returning them
- When evaluating whether to add more agents or communication rounds to an existing system
Key Technique
The Entropy Judger reframes multi-agent solution selection as a supervised classification problem: given entropy traces from an agent run, predict whether the output is correct. For each candidate solution in a pass@k set, extract a feature vector of entropy statistics across tokens, reasoning trajectories, and communication rounds. An ensemble of XGBoost and LightGBM classifiers scores each candidate f(x) in [0,1], and the candidate with the highest predicted correctness probability is selected: best = argmax over k candidates of f(x_k).
The method rests on three empirically validated observations. Certainty Preference: reducing entropy at any stage for any agent correlates with correct solutions -- high peak entropy is universally harmful. Base Uncertainty: the foundation model's intrinsic entropy during problem-solving directly bounds multi-agent performance -- if the base model is highly uncertain (entropy > 100 nats across the sequence), adding agents rarely helps. Task Awareness: entropy dynamics play different roles across task types -- math tasks show strong entropy-correctness correlation (rho up to -0.92), while code generation tasks show weaker but still useful signal.
The practical implication is that first-round entropy dominates prediction. Round-1 maximum agent entropy is consistently among the top predictive features. Extended deliberation (5+ rounds) rarely overcomes poor initial uncertainty. This means you can make go/no-go decisions about an agent run very early.
Step-by-Step Workflow
-
Define the agent topology. Choose from: single agent, sequential (planner -> solver -> critic -> judger), centralized (domain agents + coordinator), decentralized (agents with feedback loops), full-decentralized (complete graph), debate (multi-agent voting), or hybrid (two-layer with context propagation). Default to single agent unless the task demonstrably benefits from decomposition.
-
Instrument entropy collection. For each LLM call in the pipeline, capture the raw logits or log-probabilities for every generated token. Compute per-token entropy as H(t) = -sum(p(v) * log(p(v))) over the vocabulary dimension. Store these as arrays alongside the generated text.
-
Compute trajectory-level features. For each agent's response, aggregate token entropies into: total_entropy, mean_entropy, max_entropy, min_entropy, median_entropy, std_entropy, q1/q3 entropy, coefficient of variation, IQR, skewness, and tail weight. This yields ~13 core features per agent response.
-
Compute round-level transition features. Across communication rounds, track: first-to-last round entropy difference, entropy slope (linear trend), volatility (std of per-round means), and per-round agent spread (max - min entropy across agents in a round). Focus especially on round-1 metrics -- they carry the most predictive signal.
-
Extract base-model reference entropy. Run the same prompt through a single agent with the same base model. Compute its entropy features as a baseline. Calculate deltas and ratios: delta_mean = mas_mean_entropy - base_mean_entropy, ratio_total = mas_total_entropy / base_total_entropy. These ~17 features capture whether the multi-agent setup is increasing or decreasing uncertainty relative to the base model.
-
Build the feature vector. Concatenate sample-level (~13), shape (~7), round-agent (~100+), cross-round dynamics (~20), answer-token (~12), agent-level (~10), base-model comparison (~17), and computational metrics (~15 for time/tokens) into a single vector per candidate solution. The full set is 245 features.
-
Train or apply the Entropy Judger classifier. If you have labeled data (correct/incorrect for past runs), train an XGBoost + LightGBM ensemble on the feature vectors. Use 5-fold cross-validation. If no labeled data is available, use a heuristic: select the candidate with the lowest mean entropy and lowest round-1 max entropy -- the Certainty Preference principle.
-
Select from pass@k candidates. Run the multi-agent system k times on the same input. Score each candidate with the trained classifier (or heuristic). Return the candidate with highest predicted correctness: .
Concrete Examples
Example 1: Selecting the best answer from multiple agent debate runs
User: I'm running a 3-agent debate system on math problems and generating
5 candidate answers per problem. How do I pick the best one?
Approach:
1. For each of the 5 runs, collect per-token entropy from every agent
at every debate round.
2. Compute per-run feature vectors:
- mean_entropy across all agents and rounds
- max_entropy in round 1 (strongest single predictor)
- entropy_slope across rounds (is uncertainty decreasing?)
- std_entropy across agents in final round (consensus measure)
3. Rank candidates by: lowest round-1 max entropy first, then by
most negative entropy slope (uncertainty decreased most).
4. Return the top-ranked candidate.
Output (selection logic in Python):
```python
import numpy as np
def select_best_candidate(candidates):
"""Select best from pass@k using entropy heuristic.
Each candidate is a dict with:
- 'answer': str
- 'round_entropies': list of list of float
(rounds x agents, each entry is mean token entropy)
"""
scores = []
for c in candidates:
r1_max = max(c['round_entropies'][0]) # round-1 max agent entropy
all_means = [np.mean(r) for r in c['round_entropies']]
slope = all_means[-1] - all_means[0] # negative = uncertainty decreased
final_std = np.std(c['round_entropies'][-1]) # agent consensus
# Lower is better for all three components
score = r1_max + 0.5 * slope + 0.3 * final_std
scores.append(score)
return candidates[np.argmin(scores)]['answer']
Example 2: Deciding whether to use multi-agent or single-agent
User: I have a code generation task. Should I use a multi-agent system
or just a single agent?
Approach:
1. Run 20 sample problems through a single agent, collecting entropy.
2. Compute the base model's average mean_entropy across samples.
3. If mean_entropy > 80 nats: the base model is highly uncertain.
Multi-agent is unlikely to help (Base Uncertainty principle).
Invest in a better base model or better prompting first.
4. If mean_entropy is moderate (30-80): multi-agent may help.
Try a sequential (planner -> coder -> reviewer) topology.
5. If mean_entropy < 30: the model is already confident.
Single agent is likely sufficient. Multi-agent adds latency
without meaningful accuracy gain.
Decision framework:
```python
def should_use_multi_agent(base_entropies: list[float]) -> str:
avg = np.mean(base_entropies)
if avg > 80:
return "single-agent (base model too uncertain for MAS benefit)"
elif avg > 30:
return "multi-agent (moderate uncertainty, MAS can help)"
else:
return "single-agent (model already confident)"
Example 3: Early-exit from a failing multi-agent pipeline
User: My sequential agent pipeline (planner -> solver -> critic -> judger)
sometimes wastes tokens on problems it gets wrong anyway. Can I detect
failure early?
Approach:
1. After the planner (round 1) generates its output, compute its
mean token entropy and max token entropy.
2. Compare against historical thresholds from your labeled data:
- If round-1 max_entropy > threshold_high: abort and return
single-agent fallback (saves 3 agent calls).
- If round-1 mean_entropy < threshold_low: skip critic/judger,
the solver's answer is likely already correct.
3. Only continue the full pipeline for moderate-entropy cases
where multi-agent deliberation adds value.
Implementation:
```python
def early_exit_check(round1_entropies: list[float],
high_threshold: float = 120.0,
low_threshold: float = 15.0) -> str:
max_ent = max(round1_entropies)
mean_ent = np.mean(round1_entropies)
if max_ent > high_threshold:
return "abort" # Fall back to single agent
elif mean_ent < low_threshold:
return "accept" # Skip remaining agents, answer is confident
else:
return "continue" # Proceed with full pipeline
Best Practices
- Do measure entropy from the actual token logits/log-probs, not from proxy signals like response length or verbosity. Entropy computed from the probability distribution over the vocabulary is the true uncertainty signal.
- Do weight round-1 features most heavily in any selection or diagnostic scheme. The paper shows first-round entropy is consistently the top predictor across all topologies and tasks.
- Do always benchmark against a single-agent baseline before deploying a multi-agent system. In 43% of cases, a single agent wins -- you need evidence that your multi-agent setup beats it.
- Do use task-specific entropy thresholds. Math reasoning tasks show tight entropy-correctness correlation (rho ~ -0.92), while code generation is weaker. Calibrate thresholds per task domain.
- Avoid adding more communication rounds hoping to "converge" when round-1 entropy is already high. The paper shows extended deliberation rarely recovers from poor initial uncertainty.
- Avoid using multi-agent systems with base models that show very high intrinsic entropy (>100 nats average). Fix the base model or prompting first -- multi-agent amplifies base model quality, it does not compensate for it.
Error Handling
- No logit access: If using a closed API (e.g., GPT-4, Claude) that doesn't expose token logits, approximate entropy using: (a) log-probabilities if available via the API, (b) self-consistency across k samples as a proxy for uncertainty, or (c) ask the model to self-rate confidence and treat low-confidence as high entropy. The heuristic degrades gracefully.
- Feature extraction fails for short responses: Very short outputs (< 5 tokens) produce unstable entropy statistics. Pad the feature vector with the sample mean or flag these as low-confidence and default to the single-agent answer.
- Classifier overfits to one task domain: The Entropy Judger should be trained per-task or per-domain. A classifier trained on math problems will not transfer well to code generation. Use the Task Awareness principle -- always retrain or recalibrate when switching domains.
- Inconsistent entropy scales across models: Different LLMs have different vocabulary sizes, which affects raw entropy magnitude. Normalize entropy features by
log(vocab_size) to make them comparable across models.
Limitations
- Requires access to token-level logits or log-probabilities. Many commercial APIs do not expose these, limiting the technique to open-weight models or APIs with logprob support.
- The supervised Entropy Judger (XGBoost + LightGBM ensemble) requires labeled training data -- you need a set of (input, agent_run, correct/incorrect) examples to train on. The heuristic fallback (lowest entropy wins) works without labels but is less accurate.
- Entropy is a necessary but not sufficient signal. A model can be confidently wrong (low entropy, incorrect answer). The technique improves selection accuracy, it does not guarantee correctness.
- The 245-feature set was validated on math (GSM8K, AIME, MATH-500), reasoning (MMLU), and code (HumanEval) benchmarks. Performance on creative, open-ended, or subjective tasks is unvalidated.
- The approach optimizes selection from existing candidates. It does not improve the generation quality of any individual agent -- it picks the best from what is already produced.
Reference
Zhao, Y., Chen, S., & Su, N. (2026). "On the Uncertainty of Large Language Model-Based Multi-Agent Systems." arXiv:2602.04234v2. https://arxiv.org/abs/2602.04234v2
Look for: Section 4 (Entropy Judger algorithm and feature taxonomy), Table 1 (cross-validation accuracy by feature group), and Section 3 (the three key observations on Certainty Preference, Base Uncertainty, and Task Awareness). Source code: https://github.com/AgenticFinLab/multiagent-entropy