iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
LLM agents often repeat themselves because their prompts, memories, and conversations keep steering them toward the same parts of the idea space. Adding more agents or asking for longer discussion does not automatically fix that: agents can converge on one another’s proposals, while an overloaded history can keep recycling earlier answers. To get a more varied set, generate proposals independently, cue genuinely different perspectives, prune irrelevant context, and evaluate diversity separately from usefulness and feasibility.
Why do LLM agents keep producing similar ideas?
An agent’s output is shaped not only by the task, but also by what it has already generated and what the conversation keeps emphasizing. Several mechanisms can narrow its search.
Repeated reflection can feed the same ideas back in
In an iterative loop, an agent may be asked to reflect on an answer and revise it. If each reflection largely restates the previous one, the next round receives redundant input rather than a genuinely new direction. The ICLR 2025 paper Enhancing Language Model Agents using Diversity of Thoughts identifies this pattern as a limit on exploration of the decision space. It proposes diverse reflections and task-agnostic memory for retrieving lessons from earlier tasks. Its reported gains—up to 10% in Pass@1 across programming benchmarks and 13% on Game of 24 when the diverse-reflection module was integrated with Tree of Thoughts—are benchmark results, not evidence of equivalent gains in creative ideation.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Conversation history can become an anchor
Long histories can preserve useful requirements, but they can also make previous suggestions disproportionately influential. The Findings of EMNLP 2025 paper Exploring and Controlling Diversity in LLM-Agent Conversation reports that dialogue diversity degraded in long-term simulations. In its analysis, reducing contextual information increased diversity, memory had the greatest influence among the prompt components studied, and content receiving high attention consistently suppressed diversity. The paper presents Adaptive Prompt Pruning as a way to trim prompt segments. The practical challenge is to remove stale or irrelevant material without losing constraints that make an answer fit the task.
#1 Best Overall
Agent groups can converge instead of exploring
When agents see one another’s answers early, later contributions may become variations on the first proposals rather than independent searches. The Findings of ACL 2026 paper Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation examines diversity at model, cognition, and system levels. It reports authority-driven suppression of diversity, faster premature convergence under dense communication, and diminishing marginal diversity as group size increases in its studied setting. The implication is not that collaboration is inherently harmful, but that communication structure and authority can contract exploration.
Results differ across setups. A SIGDIAL 2025 study, Exploring the Design of Multi-Agent LLM Dialogues for Research Ideation, reports that larger cohorts, deeper interaction, and more heterogeneous personas enriched diversity in its research-ideation dialogues. It also found that diverse critics in an ideation–critique–revision loop improved final-proposal feasibility. These findings can coexist: more interaction or participants may help in one design, while dense coupling or scale may create diminishing returns in another. The useful variables to test are when proposals are shared, how different the roles are, how tightly agents communicate, and how many dialogue rounds the task needs.
How can you make an agent generate a more varied set?
Change the search conditions rather than simply asking for “more creative” answers. The following workflow applies the research findings while keeping task relevance in view.
- Write down the task and its non-negotiable constraints. State the audience, goal, resources, limits, and what would make an idea unusable. Preserve these requirements when trimming context; diversity without task fit is not an improvement.
- Collect independent first-pass proposals. If using multiple agents or roles, have each produce an initial proposal before showing it the other proposals. This is a practical design inference from ACL 2026 findings on dense communication and convergence, not a universal guarantee.
- Give each perspective a distinct, ordinary cue. Use roles that foreground different needs or knowledge—such as a skeptical operations lead, a first-time user, or a maintenance technician. These are illustrative prompt choices, not personas verified as tested in the cited study. The February 2026 Columbia Business School summary of Examining and Addressing Barriers to Diversity in LLM-Generated Ideas reports that ordinary personas acted as diverse sampling cues, whereas relying on a “creative genius” persona such as Steve Jobs was not the key intervention.
- Ask for a brief structured approach before ideas. Request a concise plan that explores different angles—for example, user needs, operational constraints, and unconventional alternatives—then ask for candidates. The Columbia summary reports that chain-of-thought prompting reduced fixation in its studies and that combining it with ordinary personas produced the highest idea diversity there. A short decomposition is a practical way to cue exploration; more reasoning is not guaranteed to improve originality.
- Trim context when repetition appears. Keep the brief, essential constraints, and genuinely useful discoveries. Remove obsolete suggestions, repeated reflections, and unrelated history that may anchor another round. Treat pruning as a controlled adjustment: context can also improve grounding, and the EMNLP findings do not mean less context is always better.
- Critique only after independent proposals exist. Let a differently oriented critic identify feasibility issues, then invite revisions that address those issues without copying another proposal’s core. SIGDIAL 2025 supports critic diversity in its research-ideation setup; ACL 2026 cautions that dense coupling can accelerate convergence.
- Deduplicate and assess the batch. Group near-duplicates, then evaluate the remaining ideas for novelty, feasibility, clarity, and task fulfillment as separate qualities. Retain human review for consequential decisions; an embedding filter or model-based judge is not proof that an idea is genuinely original.
- Run a small ablation. Compare the current baseline with one change at a time—such as independent drafts, role variation, or context pruning—then test the combination that appears useful. Track batch-level variety alongside human-rated usefulness and feasibility. This is a practical evaluation procedure, not a reported experiment from the cited papers.
A prompt pattern to adapt
For a single model asked to take multiple passes, try a prompt such as:
“We need ideas for [goal] for [audience]. Constraints: [essential requirements]. First, briefly map three different angles on the problem. Then propose five options that differ in their underlying approach, not just their wording. For each, state the core mechanism and one likely drawback. Do not reuse the same premise across options.”
For a multi-agent workflow, assign a different grounded perspective to each agent and ask for independent first drafts before sharing them. Avoid treating a persona label as a guarantee of different knowledge; inspect whether the resulting proposals actually differ.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should you measure whether the ideas improved?
Decide what “better” means before changing prompts. Diversity describes variation among the outputs in a set; novelty concerns how new an idea seems relative to a reference or evaluator. Feasibility asks whether it could work, while task fulfillment asks whether it addresses the brief. These dimensions can move in different directions.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →| Measure | Question it answers | Practical check |
|---|---|---|
| Diversity | How much do the proposals differ from one another? | Cluster or compare ideas for semantic overlap; inspect whether they use distinct premises or mechanisms. |
| Novelty | How new does an idea appear relative to a reference or evaluator? | Use a defined comparison set or human review; state what the ideas are being compared against. |
| Feasibility | Could the proposal work under the actual constraints? | Have an informed reviewer identify assumptions, resources, and likely failure points. |
| Task fulfillment | Does the proposal answer the brief? | Check it against the stated goal, audience, and requirements. |
A 2025 ICLR controlled study involving more than 100 NLP researchers found that LLM-generated ideas were judged more novel than expert ideas at p < 0.05, but slightly weaker on feasibility. The abstract reports no effect size, so this result should not be generalized into a claim that AI ideas are always more novel or equally workable.
Best Value
The ACL 2026 paper Automated Creativity Evaluation of Language Models Across Open-Ended Tasks introduces semantic entropy as a reference-free measure of divergent creativity, validated against human annotations and other measures, and a retrieval-based multi-agent judge for task fulfillment. It evaluates research ideation, problem solving, and creative writing, and reports that model size, temperature, recency, and reasoning can affect creative performance. The authors also report over 60% improved efficiency for their task-fulfillment evaluation framework; that figure concerns the framework’s evaluation efficiency, not creativity or idea quality. These methods are proposed evaluation tools, not universal ground truth.
The SIGDIAL 2025 study uses a Non-Duplicate Ratio based on embedding-based deduplication as a batch-level signal. Such a ratio can show whether a set contains fewer near-duplicates, but deduplication alone does not establish novelty, usefulness, or feasibility. Pair automated measures with human or task-specific assessment.
What to take from the evidence
No single prompt or agent architecture is shown to make every system reliably original. The findings point instead to controllable factors: redundant reflection, memory and context, persona cues, communication structure, and how diversity is evaluated. Test those factors on the task you actually care about, and do not optimize variety at the expense of relevance or feasibility.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
