Self-correction
自纠错
The agent critiques, retries, edits, or calls tools inside one episode. No reusable state must be committed.
Agent 在单个任务内批评、重试、修改或调用工具，不要求提交可复用状态。
retryreflectiontest-time searchScope and strict definition
范围与严格定义
The term “self-evolving” is frequently used for systems that only retry. We separate transient refinement from persistent learning.
“自进化”经常被用于仅增加重试的系统，因此必须区分临时修正与持久学习。
Self-evolution
自进化
Experience produces a persistent update to model parameters or scaffold components—prompt, memory, skills, tools, workflow, topology, reward, world model, or embodiment model—and the update is evaluated on future tasks.
经验对模型参数或 scaffold 组件产生持久更新，包括提示词、记忆、技能、工具、工作流、拓扑、奖励、世界模型或具身模型，并在未来任务上验证。
persistent updatefuture-task gaintransfer / forgettingKnowledge map
知识地图
Four orthogonal axes organize the field: update target, feedback source, timescale, and commitment mechanism.
使用四个正交维度组织该领域：更新对象、反馈来源、时间尺度与提交机制。
What evolves?
进化什么？
- Foundation-model parameters
- Prompt and context
- Memory and skills
- Tools and workflow
- World and embodiment models
Which feedback?
由什么反馈驱动？
- Scalar environment reward
- Natural-language critique
- Tool or verifier output
- Peer-agent feedback
- Counterfactual visual evidence
When?
何时进化？
- Intra-episode
- Inter-episode
- Offline iterative
- Online continual
- Open-ended lifelong
How validated?
如何验证？
- Matched compute budget
- Future-task probes
- Transfer and forgetting
- Negative evolution
- Versioning and rollback
Research branches
研究分支
Click a branch to filter the literature matrix. The taxonomy distinguishes actual committed updates.
点击分支筛选论文矩阵；分类依据是系统实际提交的更新。
Literature matrix
论文矩阵
The decisive field is the committed update—not whether a paper uses “evolving” in its title.
关键字段是“提交了什么更新”，而不是标题是否出现 evolving。
| # | Year | Paper | Venue | Branch | Committed update | Feedback |
|---|
Official proceedings are preferred where available; otherwise the entry remains marked as arXiv.
Open-problem map
开放问题地图
Mechanism-level comparison reveals ten recurring weaknesses.
机制层面对比揭示了十个反复出现的问题。
Visual critique is not a trustworthy reward
Self-generated critiques miss perceptual errors and can damage correct trajectories.
Evidence-grounded verification and counterfactual replay.
“Evolution” is often context engineering
Many systems change prompts or memory without demonstrating new model competence.
Compare scaffold-only and parameter updates explicitly.
No persistent or monotonic gain guarantee
Next-task gains can coexist with forgetting, pollution, and reward hacking.
Longitudinal probes, versioned commits, and rollback.
Visual evidence is compressed into text
Long-term memory loses decisive pixels, temporal details, and object-state changes.
Evidence-preserving visual–verbal memory.
The update target is fixed by the designer
Memory, tool, perception, and policy failures receive the same update.
Failure-conditioned multilevel routing.
Self-generated curricula collapse
Questioners drift to invalid, repetitive, narrow, or reward-exploiting tasks.
Capability-frontier sampling and skill coverage.
Embodiment is assumed stationary
Camera geometry, sensing, latency, and action dynamics change after deployment.
Active self-calibration and embodiment-aware world models.
Successful rollouts are not safety contracts
A few successful executions do not establish safe reuse in unseen conditions.
Counterexample search, contracts, and skill rollback.
Retry is conflated with learning
Larger inference budgets and persistent capability accumulation are mixed.
Matched-budget controls and fixed post-commit probes.
Cost and autonomy are underreported
Human seeds, proprietary teachers, resets, and tool calls can dominate gains.
Report labels, calls, rollouts, GPU hours, and interventions.
Idea iteration log
Idea 迭代记录
Obvious directions were rejected after literature collision checks instead of being repackaged as novelty.
直观方向在文献碰撞检查后被主动否决，而不是重新包装成新颖性。
| Version | Candidate | Decision | Reason |
|---|---|---|---|
| v0 | Self-generated visual questions + GRPO | Rejected | Direct overlap with VisPlay, Active-Zero, Agent0-VL, and RISE. |
| v1 | Reflection + persistent experience memory | Rejected as main idea | Crowded by Mobile-Agent-E, A-MEM, SEAgent, SE-VLN, and related systems. |
| v2 | Actor–critic visual self-correction | Rejected as main idea | VISCO, Critic-V, grounding correction, and perception-centric reward models cover the core. |
| v3 | Counterfactual validation of visual experience | Promising component | Makes vision indispensable and rejects spurious lessons. |
| v4 | Evidence-gated routing across update surfaces | Refine | Generic evidence verification and generic update gates are no longer sufficient novelty. |
| v5 | Visual causal evidence paths + reliance-preserving commit | Recommended | Focuses novelty on interactive counterfactual replay, update routing, hidden-grounding protection, and rollback. |
CVPR candidate ideas
CVPR 候选 Idea
Scores are comparative research judgments, not experimental results.
评分是相对研究判断，不是实验结果。
#1
GroundEvo
Causally-Grounded Multilevel Self-Evolution for Visual Agents
基于视觉因果证据的多层级 Agent 自进化
A visual agent should not absorb every self-generated experience. It should identify a causal evidence path across regions, frames, state changes, decisions, and actions; verify it through active re-observation and minimal counterfactual replay; and only then update memory, skills, or parameters.
视觉 Agent 不应无差别吸收所有自生成经验。它应先定位由区域、帧、状态变化、决策与动作组成的因果证据路径，通过主动重观察和最小反事实重放验证，再更新记忆、技能或参数。
Novelty
8/10
CVPR fit
9/10
Feasibility
7/10
Risk
6/10
1 · DiagnoseLocate perception, planning, memory, tool, or execution failure.
2 · Re-observeCrop, zoom, temporal replay, alternate view, and visual tools.
3 · CounterfactMinimally alter the suspected visual factor and replay.
4 · Route updateOne-off → memory; repeated → skill; systematic → parameters.
5 · Commit safelyTask probes + evidence-reliance probes + rollback.
Core method blocks
- Trajectory decomposition into visual claims, evidence, decisions, actions, and outcomes.
- Interactive evidence verification rather than free-form textual reflection.
- Environment counterfactual replay to distinguish causal from spurious lessons.
- Failure-conditioned routing among episodic memory, procedural skills, and LoRA/policy updates.
- Versioned provenance, old-task replay, evidence-reliance probes, and rollback.
Novelty boundary
Not generic evidence verification and not a generic update gate. The paper must jointly demonstrate visual causal paths, environment counterfactuals, multilevel routing, and protection against hidden grounding forgetting.
#2
EgoShift
Self-Calibrating Embodied Agents under Sensor and Embodiment Drift
An embodied agent actively diagnoses changes in camera geometry, sensing, latency, or action dynamics, then evolves its perception–action calibration and world model.
Novelty
9/10
CVPR fit
10/10
Feasibility
5/10
Risk
8/10
Method blocks
- Prediction-residual drift detector.
- Agent-generated low-risk diagnostic micro-actions.
- Online camera/action calibration and world-model update.
- Safety probes before committing adaptation.
Assessment
Strongest visual and embodiment identity, but substantially higher engineering cost.
#3
ViMEvo
Self-Evolving Visual Memory with Evidence Fidelity
Learn which crops, frames, object states, and textual abstractions must be preserved, revised, and retrieved under a fixed memory budget.
Novelty
7/10
CVPR fit
9/10
Feasibility
9/10
Risk
4/10
Method blocks
- Dual visual–verbal memory with object-state links.
- Visual necessity and evidence-fidelity estimators.
- Stale-memory revision after state changes.
- Budgeted consolidation trained from downstream action reward.
Assessment
Best low-risk demo, but it must target decisive visual evidence rather than another vector database.
#4
MetaEvolve-V
Failure-Conditioned Routing across Memory, Tools, Workflow, and Parameters
A meta-controller diagnoses the failure type and chooses the cheapest effective evolution surface.
Novelty
8/10
CVPR fit
7/10
Feasibility
5/10
Risk
8/10
Method blocks
- Visual failure taxonomy and causal diagnosis.
- Actions over prompt, memory, tool, workflow, and LoRA updates.
- Cost-aware contextual bandit or meta-RL router.
- Rollback on regression or incompatibility.
Assessment
Conceptually broad, but it must be restricted to two or three update surfaces.
Experiment path
实验落地路径
Start with a falsifiable demo and expand only if causal verification reduces harmful learning.
先完成可证伪 Demo；只有因果验证确实减少错误学习后才扩展。
D0 · Pipeline check
- Open 7B-class VLM agent.
- One GUI environment with controlled visual perturbations.
- 50–100 tasks.
- Memory-only vs evidence-gated memory.
- Primary test: fewer wrong lessons and higher next-task success.
D1 · Controlled study
- Layout, scale, theme, occlusion, distractor, and state changes.
- Matched rollout, token, and tool-call budgets.
- Ablate re-observation, counterfactual replay, router, and rollback.
- Report transfer, forgetting, harmful updates, and cost.
D2 · CVPR evidence
- Two VLM backbones.
- At least two visual-agent domains.
- GUI/Web plus VLN or embodied manipulation.
- Cross-domain transfer and long-sequence evaluation.
- Human audit of evidence paths.
| Capability | Metric | Why |
|---|---|---|
| Immediate adaptation | Post-update success / SPL / grounding accuracy | Does the update help? |
| Persistent gain | Fixed-probe score after later tasks | Is it genuine accumulation? |
| Transfer | Unseen-task and unseen-environment gain | Was a reusable lesson learned? |
| Forgetting | Backward transfer / worst-task drop | Was previous competence erased? |
| Hidden grounding drift | Counterfactual evidence-reliance score | Does the agent still use correct evidence? |
| Negative evolution | Fraction of harmful committed updates | Is the gate reliable? |
| Efficiency | Rollouts, calls, tokens, GPU hours, memory size | Are gains budget-matched? |
Recommended first paper question
推荐先验证的论文问题
Do self-evolving visual agents learn spurious lessons from successful and failed trajectories, and can interactive visual counterfactual verification reduce harmful updates while preserving transferable improvement?
自进化视觉 Agent 是否会从成功与失败轨迹中学到伪相关经验？交互式视觉反事实验证能否在保留可迁移收益的同时减少有害更新？
Novelty
- Not another questioner–solver loop.
- Not reflection memory only.
- Explicit 2026 collision analysis.
Vision centrality
- Text-only ablation must fail.
- Region/frame/state evidence is necessary.
- Visual changes are controlled.
Evolution validity
- Persistent committed update.
- Future-task probes.
- Matched inference budget.
Reliability
- Harmful updates measured.
- Forgetting and contamination tested.
- Provenance and rollback included.