Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery

Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery | Signal Canvas | ScienceToStartup

Page Freshness

Signal Canvas proof surface

Canonical route: /signal-canvas/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery

stale

Proof freshness: stale
Proof status: unverified
Display score: 4/10
Last proof check: 2026-03-30
Score updated: 2026-04-02
Score fresh until: 2026-05-02
References: 35
Source count: 3
Coverage: 50%

This page is showing the last landed evidence receipt and score bundle because the latest proof data is outside the freshness window.

Agent Handoff

Canonical ID can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery | Route /signal-canvas/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery

REST example

curl https://sciencetostartup.com/api/v1/agent-handoff/signal-canvas/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery

MCP example

{
  "tool": "search_signal_canvas",
  "arguments": {
    "mode": "paper",
    "paper_ref": "can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery",
    "query_text": "Summarize Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery"
  }
}

source_context

{
  "surface": "signal_canvas",
  "mode": "paper",
  "query": "Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery",
  "normalized_query": "2603.26177",
  "route": "/signal-canvas/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery",
  "paper_ref": "can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery",
  "topic_slug": null,
  "benchmark_ref": null,
  "dataset_ref": null
}

Evidence Receipt

Route status: building

Claims: 7

References: 35

Proof: Verification pending

Freshness state: computing

Source paper: Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery

PDF: https://arxiv.org/pdf/2603.26177v1

Source count: 3

Coverage: 50%

Last proof check: 2026-03-30T21:57:28.809Z

Signal Canvas receipt window

Not build-ready: Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery

/buildability/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery

Ignoreblocked

Subject: Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery

Verdict

Ignore

Verdict is Ignore because current viability and proof state do not clear the buildability gate.

Preparing verified analysis

GitHub Code Pulse

No public code linked for this paper yet.

Claim map

Strong 7Mixed 0Weak 0

Evidencepartial
Access to feedback yields a +53.4% increase in discoveries per feature on average (p = 0.003).
Implicationpartial
This is a direct quantitative result stated in the abstract with a p-value.
Verificationpartial
partial
Evidencepartial
Under this control, the performance gain disappears, indicating that the observed improvement depends on the structure of the feedback signal (+13.0 hits, p = 0.003).
Implicationpartial
This is a key experimental finding presented in the abstract to validate the learning mechanism, with a p-value supporting the significance.
Verificationpartial
partial
Evidencepartial
Upgrading from Claude Sonnet 4.5 to 4.6 reduces gene hallucination rates from ~33%--45% to ~3--9%
Implicationpartial
This is a specific quantitative technical improvement reported in the abstract.
Verificationpartial
partial
Evidencepartial
converting a non-significant ICL effect (+0.8, p = 0.32) into a large and highly significant improvement (+11.0, p=0.003) for the best ICL strategy.
Implicationpartial
This is a direct comparison of model capabilities and their impact on learning, with specific quantitative results and p-values.
Verificationpartial
partial
Evidencepartial
These results suggest that effective in-context learning from experimental feedback emerges only once models reach a sufficient capability threshold.
Implicationpartial
This is a concluding statement derived from the comparison of different LLM versions and their learning performance.
Verificationpartial
partial
Evidencepartial
Both feedback-enabled LLMs outperform the GP-UCB baseline ( 19.5, p= 0.003 ).
Implicationpartial
This is a direct comparison of the proposed method against a baseline with quantitative results and statistical significance.
Verificationpartial
partial
Evidencepartial
First, the ICL-EF agent is highly efficient at ex-ploitation. Once a few hits in major protein complexes (e.g., transcription machinery, nuclear envelope, or the ubiquitin-proteasome system) are found, the basic feedback mecha-nism allows ICL-EF to
Implicationpartial
This is a qualitative explanation for the performance of a specific agent, supported by observations.
Verificationpartial
partial

Author intelligence and commercialization panels stay hidden until the proof receipt is verified, cites at least 3 references, includes at least 2 sources, and clears 50% coverage. The paper narrative and citation surfaces remain public while verification is pending.

Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery

Use Signal Canvas as the narrative proof surface

Use this Signal Canvas via API or MCP

Signal Canvas proof surface