acceptodds
Withdrawn from ICLR 2027

Beyond Context: Parameter-Efficient Model-Level Evolution of Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agents accumulate experience, and most frameworks keep it in context—editable skills, memories, and workflows that stay open to inspection and revision. That openness is worth paying for, but the payment recurs: text-carried behavior occupies the context of every call that uses it, long after it has stabilized. We organize persistent agent experience into versioned, callable domain capabilities, each encoded in a separately stored parameter-efficient update to a frozen shared backbone—a LoRA patch in our implementation. We train these patches from verified records of an agent's own work and teacher-assisted supervision; only those that pass evaluation are enabled for invocation by later agent versions. At run time the parent agent selects an enabled capability inside its own task-solving process, with no separately trained dispatch module; a deterministic harness assembles the specialist's request from the task, accepted artifacts, and runtime observations under an explicit output contract. Specialists cooperate through what they produce, sharing the backbone while keeping their patches separate; editable skills remain available for coordination and rapid revision. We evaluate a registry of eight domain patches on a Qwen3-8B backbone. Each patch specializes in its domain and improves on the base model on every evaluated benchmark that contributed no training rows. Behind the agentic interface, Ours uses patch-backed specialists with typed handoffs and outperforms the base model in every reported domain. On five constructed workflows that none of the patches was trained for, the registry grounds 274 of 500 cases against 173 on base weights, a 58% relative gain on tasks requiring two to four specialist stages to cooperate. Holding all eight patches resident on one accelerator uses 82.25% less process GPU memory than eight full-model replicas, with low runtime switching latency (32.84 ms median).

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, parameter-efficient fine-tuning, LoRA, agent evolution, modular adaptation, experience consolidation, multi-adapter serving, skill libraries

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2HJVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors train 8 domain LoRA adapters on a frozen Qwen3-8B backbone and exposes them as callable specialists to a parent agent. A deterministic harness builds specialist requests from the task, prior artifacts, and runtime observations. The proposed framework also describes evaluation-gated adapter updates and version management when the LoRA can be reconfigured and installed. Experiments report domain-specialization gains, improved success on five constructed multi-domain workflows, and low resident memory than maintaining 8 full-model replicas.

Critical strengths
  • Updating domain capabilities without retraining the shared backbone is a useful direction. The paper implements a working specialist registry and reports gains on several domain benchmarks.
Critical weaknesses
  • The technical contribution beyond existing multi-LoRA systems is limited. Punica and S-LoRA already serve multiple adapters over a shared backbone [1, 2], so the memory saving over eight full-model replicas is an expected benefit of that design. Treating LoRAs as callable tools is also explored by Adaptive Minds [3]. The remaining additions are mainly version management, evaluation gates, and task-specific handoffs. Yet the experiments evaluate a fixed adapter library rather than showing successive experience-driven updates or an advantage over ordinary offline fine-tuning. I do not see a substantial new system mechanism or an effective agent-evolution method established here.

  • The evaluation is limited to prepared tasks and leaves new-capability validation unresolved. All adapters use one 8B backbone, and multi-domain evaluation consists of five constructed workflows with task-specific verifiers and handoff contracts. These results show that the adapters work within the supplied setups, but say little about broader agent tasks. More fundamentally, I could not see how the system defines a new capability, verifies its outputs, or decides whether an update is ready to deploy when a suitable evaluation harness is not already available. If each addition requires the user to build that harness and specify the acceptance criteria, much of the claimed evolution remains manual task engineering. The current evaluation does not demonstrate a complete process for acquiring and validating new capabilities.

[1] Lequn Chen et al. Punica: Multi-Tenant LoRA Serving. MLSys 2024. https://arxiv.org/abs/2310.18547

[2] Ying Sheng et al. S-LoRA: Serving Thousands of Concurrent LoRA Adapters. MLSys 2024. https://arxiv.org/abs/2311.03285

[3] Pavan C. Shekar and Aswanth Krishnan. Adaptive Minds: Empowering Agents with LoRAs as Tools, 2025; v2, June 2026. https://arxiv.org/abs/2510.15416v2

Questions for authors
  • What concrete capability is new beyond shared-backbone multi-LoRA serving and LoRA-as-tools orchestration? What successive experience-to-update cycles were actually evaluated? Please report incumbent-versus-candidate results, gate thresholds, accepted and rejected updates, and the resulting agent performance, compared with ordinary offline training using the same data.
  • When a new task or capability is introduced, who defines its scope, verifier, evaluation tasks, and acceptance criteria? Can the authors show a complete example of this process and specify the manual work required, rather than starting from a prepared domain-specific harness?
Reviewer_BweTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an approach using domain-specific LoRA patches over a frozen model backbone to adapt agents, specifically a base agent that invokes specialist patches and combines their outputs. They evaluate specialisation across eight domains (SQL, finance, medicine, cybersecurity, law, chemistry, coding, and data analysis) and show that their adapted model can, on average, match or improve over an unadapted large generalist model baseline.

Critical strengths
  • The range of domains evaluated by the paper is comprehensive and provides useful coverage of agent specialisation beyond a single task.
  • The separation between different internal baselines (Pure, Concat, Query) is competent and clearly shows the effect of each setup on the performance.
Critical weaknesses
  • My most serious concern is with the novelty of this work. The main ingredients are all familiar: supervised specialisation with LoRA, libraries of adapters, agent-orchestrated invocation of specialist. It's unclear to me what makes a "patch" substantively different from a parameterised skill aside from its invocation interface, metadata, and management procedures.

  • The paper's results are on Qwen3-8B and Qwen3-235B-A22B-Instruct-2507, both of which are very old by now (having been released in ~mid-2025). Both models have direct successors of a roughly similar parameter size in Qwen3.5-9B and Qwen3.5-122B-A10B-FP8. Moreover, the paper explicitly does not invoke thinking for these models. My concern is therefore the limited usefulness of the paper's results to the wider research community, as this does not typically reflect practical usage of these models as agents.

  • The Skill baseline adds only 128–209 tokens of instructions, which seems to me like a weak baseline. This is important because a core motivation of the paper is that context is costly. Compare S2L, for example, which looked at full SKILL.md documents, or context-evolution methods (ACE, AWM, ExpeL) which the paper frames itself against. As is stands the paper's central weights-versus-context claim has not been tested against a real context-evolution method.

Questions for authors
  • Could you include more recent models, and thinking models, in the paper's evaluation?

  • More importantly, could you include existing literature method baselines such as context-evolution methods (like ACE, AWM, ExpeL), adapter selection methods (like LoraHub or LoraRetriever), single multi-task LoRA adapters, or even a merged adapter over all of your eight domains? I would be happy to revise my score upwards if the main methodology holds against a single multi-task LoRA plus one context-evolution method.

  • I noticed that the Qwen 235B’s results varies a lot across workflows (28 on law versus 82 on finance). Might this suggest sensitivity to prompt and parser format?

Reviewer_uAVLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper serves LLM with multiple Lora. Update the Loras during online agent execution. It trains these lora patches from verified records of an agent's own work and teacher-assisted supervision; only those that pass evaluation are enabled for invocation by later agent versions. At run time the parent agent selects an enabled capability inside its own task-solving process, with no separately trained dispatch module; a deterministic harness assembles the specialist's request from the task, accepted artifacts, and runtime observations under an explicit output contract.

On five constructed workflows that none of the patches was trained for, the registry grounds 274 of 500 cases against 173 on base weights, a 58% relative gain on tasks requiring two to four specialist stages to cooperate.

Critical strengths
  1. Using multiple Loras for agent serving instead of skills evolution is an interesting topic.

  2. Routing to task-specific lora is an effective method without sharing the full context. Like subagent mode.

  3. Experiments on eight specialists improve performance across multiple benchmarks and reduce GPU memory usage compared to a big model.

Critical weaknesses
  1. Training separate LoRA adapters and selecting them during inference has been studied a few years. A few search [1-4].

  2. The paper does not compare against merging all LoRAs or jointly fine-tuning a single multi-domain LoRA with all data.

  3. The self-evolution is not supported by experiments. The lora is only trained once and multi-round cycle can be helpful.

[1] Adaptive Minds: Empowering Agents with LoRA-as-Tools. 2025

[2] Towards Modular LLMs by Building and Reusing a Library of LoRAs. 2024

[3] LoRA Retriever: Input-Aware LoRA Retrieval and Composition for Mixed Tasks in the Wild. 2024

[4] LoRAHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition. 2023

Questions for authors

As in weaknesses. The main concern is on comparing with a single multi-domain LoRA and the self-evolution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.