---
title: 'HARPA: Dual Frameworks in Discovery & Seismology'
url: https://www.emergentmind.com/topics/harpa
type: topic
---

# HARPA: Dual Frameworks in Discovery & Seismology

Searching arXiv for HARPA-related papers to ground the article in current sources.
arxiv_search(query="HARPA", max_results=10)
arxiv_search(query="HARPA automated scientific discovery", max_results=10)
HARPA is an acronym used on arXiv for two unrelated technical systems with distinct disciplinary meanings. In automated scientific discovery, HARPA denotes “HARPA: A Testability-Driven, Literature-Grounded Framework for Research Ideation,” a framework for generating hypothesis-driven research proposals that are grounded in the literature, testable by an ASD agent, and adaptable to prior experimental outcomes [2510.00620]. In seismology, HARPA denotes “HARPA: High-Rate Phase Association with Travel Time Neural Fields,” a framework for associating seismic phases in dense, high-confusion regimes while jointly inferring heterogeneous wave speeds [2307.07572]. The acronym therefore has no single domain-independent definition; its meaning is determined by research context.

## 1. Disambiguation and nomenclature

The two arXiv uses of HARPA can be distinguished by domain, expansion, and technical objective.

| Domain | Expansion | Core objective |
|---|---|---|
| Automated scientific discovery | Hypothesis And Research Proposal Assistant | Generate literature-grounded, testable research proposals for ASD agents |
| Seismology | High-Rate Phase Association | Associate interleaved seismic arrivals under unknown heterogeneous wave speeds |

Confusion between these systems is a recurrent risk because both are presented as full-stack frameworks rather than narrow modules. The research-ideation HARPA starts from a “source paper,” mines related literature, explores a hypothesis design space, and learns from execution outcomes [2510.00620]. The seismological HARPA starts from station picks and infers event parameters and a latent velocity code by minimizing an optimal-transport misfit through a travel time neural field [2307.07572]. A common misconception is therefore to treat HARPA as a single methodology; the available arXiv record instead supports only a shared acronym.

## 2. HARPA as a framework for research ideation

In automated scientific discovery, HARPA was introduced to address two linked deficiencies of existing ideation tools: they often generate hypotheses that are not simultaneously testable and grounded in the scientific literature, and they are not adaptive to prior experimental outcomes [2510.00620]. Its design is explicitly modeled on a human research workflow in which ideation is multi-stage and iterative: mining trends, exploring a structured hypothesis space, and converging on specific research gaps.

The proposal generator begins from a user-provided “source paper” and proceeds through three phases. First, in literature mining and trend identification, it constructs a temporal “reasoning chain” of related papers via citation graph traversal and identifies an initial generic “seed hypothesis” \(H\) reflecting an emerging research gap. Second, in hypothesis design space exploration, it performs Socratic question answering and variable extraction. The question-generation step creates approximately 20 targeted questions,
\[
\mathcal{Q} = \mathcal{F}_{QA}(H, Agent_{desc}),
\]
which are answered over retrieved literature snippets \(RW\) to refine \(H\) into a more specific \(H'\). It then extracts key variables,
\[
KV = \mathcal{F}_{extract\_var}(H'),
\]
and retrieves possible values and metadata from the literature to construct
\[
\mathcal{V}_{space} = \mathcal{F}_{extract\_space}(H', KV, RW).
\]
Third, in gap pinpointing and hypothesis convergence, it samples final hypotheses
\[
H_{\text{final}} = \mathcal{F}_{generate}(H', \mathcal{V}_{space}, RW),
\]
with the dual requirement that they combine variable assignments in novel ways and fill an under-explored gap.

Each final hypothesis is packaged as a proposal artifact containing rationale, related work, variable definitions, an implementation plan through CodeScientist’s “idea\(\rightarrow\)implementation” module, and references. The methodological emphasis is therefore not only on ideation but on executable ideation: proposals are intended to be actionable under the capabilities and limits of an ASD agent rather than merely conceptually interesting.

## 3. Literature grounding, enumeration, and adaptive scoring

The framework’s literature-grounding layer uses the S2ORC snippet index, iterative retrieval queries formed by generalizing the seed hypothesis into multiple abstraction levels, and a citation graph that chains papers that “scientifically depend” on the source via citing relations [2510.00620]. All language-model calls used in Socratic QA, refinement, variable extraction, and final generation are performed with GPT-4o via a function-calling API. This establishes a retrieval-centered architecture in which grounding is not post hoc justification but an explicit input to every stage of proposal construction.

Candidate proposals are enumerated by sampling combinations from \(\mathcal{V}_{space}\). A lightweight untrained LLM “baseline scorer” can rank proposal pairs, but HARPA replaces this with a dedicated reward model trained to predict which proposal is more likely to execute successfully on a given ASD agent. The scorer is trained under the RM-R1 framework in two stages. During data generation, each proposal \(P\) is executed with CodeScientist to obtain a raw trace \(E\), and the trace is assigned a categorical label:
\[
\text{Label}(E) =
\begin{cases}
\text{Success}, & \text{if } faithfulness\_category = \text{faithful},\\
\text{Failure}, & \text{if } faithfulness\_category = \text{errors} \;\vee\; (faithfulness\_category=\text{inconclusive}\;\wedge\;hypothesis\_category=\text{inconclusive}),\\
\text{Uncertain}, & \text{otherwise.}
\end{cases}
\]
Pairwise preferences \(\bigl(P_a \succ P_b\bigr)\) are then formed when one proposal “wins” over another, and an oracle LLM such as Claude-Sonnet-4 produces a rubric-style reasoning trace.

The training phase consists of supervised distillation followed by preference optimization. In the first stage, Qwen-7B-Instruct is fine-tuned on oracle rationales so that the student model reproduces rubric-style explanations. In the second, PPO-based RLVR aligns the scorer to “Success vs. Failure” labels. A notable design choice is conditioning on the ASD agent profile: compute limits, dataset access, and human-in-loop constraints are concatenated to the input so that the scorer adapts to different execution environments. The final model outputs both a pairwise judgment and an interpretable reasoning trace that highlights execution success, complexity, and cost efficiency. This suggests that HARPA treats feasibility as an environment-relative property rather than an absolute one.

## 4. Empirical evaluation and implementation profile of the ideation framework

The human-centric evaluation compares HARPA with AI-Researcher. Twelve domain experts each provided 1–2 source papers and rated 2 HARPA and 2 baseline proposals, for 40 proposals per system, on a 10-point Likert scale across Familiarity, Novelty, Feasibility, Expected Effectiveness, Excitement, Overall, Confidence, and four hypothesis-specific dimensions: Literature Grounding, Motivation, Coherence, and Specificity [2510.00620]. The reported mean differences \( \text{HARPA} - \text{baseline} \) show significant gains in Feasibility, \(+0.78\) \((p<0.05,\ \text{bootstrap})\), and Literature Grounding, \(+0.85\) \((p<0.01)\). Specificity, Motivation, and Overall show positive trends without statistical significance. Novelty shows a modest decrement, with HARPA at 5.98 versus the baseline at 6.43; the paper interprets this as indicating that HARPA’s ideas remain sufficiently novel.

The ASD-centric execution study links expert judgment to actual automation outcomes. HARPA generated 1,222 proposals over 275 ACL papers, each executed five times in CodeScientist, yielding 1,222 outcome labels. Among the 40 user-evaluated proposals, HARPA produced 20 successful executions versus 11 for the baseline and 16 failures versus 21 for the baseline. The per-idea success rate is reported as \(0.24\) for HARPA versus \(0.065\) for the baseline \((p<0.001,\ \text{bootstrap})\). The paper summarizes this as HARPA proposals succeeding approximately \(2\times\) more often in automated experiments.

The scorer evaluation reports held-out pairwise accuracy of 0.52 for the untrained LLM baseline on \(N=186\) success/failure pairs and 0.81 for HARPA-Scorer, corresponding to a \(+0.28\) absolute gain and a 53% relative gain. Pairwise rubric consistency with the oracle is 0.80 for Execution Success, 0.79 for Faithfulness, 0.67 for Complexity, 0.73 for Cost Efficiency, 0.69 for Hypothesis Validity, and 0.55 for Interestingness. Rationale overlap against reference traces is also higher, with BLEU 0.22 versus 0.08 and ROUGE-1/2/L of 0.55/0.22/0.26 versus 0.43/0.13/0.18.

The reported implementation stack is likewise specific. The generator uses GPT-4o, a Python pipeline, and S2ORC snippet search. The scorer uses a Qwen-7B-Instruct backbone together with OpenRLHF, DeepSpeed, and verl’s PPO trainer. Training hardware is NVIDIA A100-SXM4-80GB GPUs, with \(4\times\) for distillation and \(8\times\) for RL. The paper identifies three operational constraints: dependence on snippet retrieval and citation-graph quality, reward-model training solely on CodeScientist traces, and the time cost of the multi-stage pipeline. Proposed extensions include richer execution environments such as wet-lab or human studies, multi-agent ideation, active learning for high-information-gain experiments, and end-to-end “robot scientist” demonstrations. The code, pretrained scorer, and ASD execution trace dataset are stated to be publicly released.

## 5. HARPA as high-rate phase association in seismology

In seismology, HARPA addresses a different problem: associating seismic arrivals with their originating earthquakes when event rates are high enough that arrivals from different events interleave across stations and the ordering of arrivals at each station no longer reflects event origin order [2307.07572]. The target setting is dense seismicity, including microseismic clouds and very small earthquakes obeying the Gutenberg–Richter law. Traditional association methods either assume a simple velocity model or depend on well-separated events so that picks can be grouped by time windows and associated event-by-event through fixed-velocity travel-time tables.

The framework is motivated by two failure modes of such methods in the “high-rate” regime. First, arrival sequences at different stations become almost uncorrelated, expressed as a high confusion factor with \(CF \to 1\). Second, the true three-dimensional wave-speed structure can no longer be neglected, yet is typically unknown or only coarsely approximated. Under these conditions, fixed-velocity event-by-event methods, including back-projection, Bayesian GMMs, and PhaseLink, break down because travel-time residuals become too large or because labeled training sets covering all heterogeneous velocity structures are infeasible.

HARPA responds by combining three ideas. It jointly estimates earthquake source parameters and the underlying velocity model in a low-dimensional latent space. It replaces combinatorial phase assignment with measure matching based on optimal transport, specifically Wasserstein distances between observed and predicted arrival-time measures. It also introduces a continuous differentiable travel time neural field that maps source location and latent velocity code to arrival times at each station, enabling end-to-end gradient-based and sampling-enhanced inference. The result is a formulation in which phase association is recast as a continuous inverse problem rather than an explicit integer assignment problem.

## 6. Mathematical formulation, architecture, and empirical behavior in seismology

The formal problem is posed over a volume of interest \(\Omega \subset \mathbb{R}^3\) with stations \(\partial \Omega\) at known receiver locations \(\{r_i\}_{i=1}^N\). Over a fixed time window there are \(M\) events with unknown locations and origin times,
\[
e_j = (s_j,\tau_j), \quad j=1,\dots,M,
\]
and station \(i\) records picks
\[
\mathcal{D}_i = \{t_{i,k} \mid k=1,\dots,K_i\}.
\]
With a known speed model \(c(x)\), each pick satisfies
\[
t_{i,k} = T_c(s_j; r_i) + \tau_j
\]
when pick \(k\) at station \(i\) originates from event \(j\). HARPA departs from the fixed-\(c\) assumption by learning a latent generative model of wave speeds and a neural approximation \(T_\theta(s,z;r_i)\) to travel times.

The velocity model is trained from realistic three-dimensional velocity fields \(\{c_n(x)\}\), each discretized on a \(32\times 32\times 32\) grid. An autoencoder \((E_\phi, D_\psi)\) embeds these fields into a latent space \(z \in \mathbb{R}^L\) with \(L \approx 4\)–6:
\[
z = E_\phi(c), \qquad c \approx D_\psi(z).
\]
For any latent code \(z\) and source-receiver pair \((s,r_i)\), travel times derived by a fast-marching method are fitted by a continuous neural field \(T_\theta\). The network is specified as a SIREN of depth approximately 6 layers and width approximately 256, with frequency parameter \(w_0 \approx 15\); inputs \((s,z,r_i)\) are normalized to \([0,1]\), and gradients with respect to \(s\) and \(z\) are obtained by automatic differentiation.

Observed and predicted arrivals are lifted to empirical probability measures. For station \(i\),
\[
\rho_i^{obs} = \frac{1}{K_i}\sum_{k=1}^{K_i}\delta_{t_{i,k}}, \qquad
\rho_i^{model}(E,z) = \frac{1}{M}\sum_{j=1}^{M}\delta_{T_\theta(s_j,z;r_i)+\tau_j}.
\]
The total misfit is
\[
W(E,z) = \sum_{i=1}^{N} w_p\!\left(\rho_i^{obs}, \rho_i^{model}(E,z)\right),
\]
where \(w_p\) is the \(p\)-Wasserstein distance. To handle spurious or missing picks, HARPA uses an unbalanced optimal transport formulated as a sparsity-constrained linear assignment. Because \(W(E,z)\) is differentiable through \(T_\theta\), gradient-based optimization is possible, but the objective is nonconvex with many poor local minima. HARPA therefore uses Stochastic Gradient Langevin Dynamics:
\[
z \leftarrow z - \eta \nabla_z W + \sqrt{2\eta}\,\epsilon\,\xi_z,
\]
\[
s_j \leftarrow s_j - \eta \nabla_{s_j}W + \sqrt{2\eta}\,\epsilon\,\xi_{s_j},
\]
\[
\tau_j \leftarrow \tau_j - \eta \nabla_{\tau_j}W + \sqrt{2\eta}\,\epsilon\,\xi_{\tau_j},
\]
with i.i.d. Gaussian noise terms \(\xi\). The paper also notes an optional Time Differences of Arrival formulation that removes \(\tau_j\).

The reported experiments are organized around synthetic overthrust models, unknown-velocity Gaussian random field experiments, robustness analyses, and ablations. In a synthetic overthrust setting with known \(c\), 8 events, 20 stations, and \(CF \approx 0.93\), HARPA achieves \(ACC=95.0\%\) phase association, average pick residual \(e_\ell=0.03\,\text{s}\), and source-location \(RMSE \approx 3\,\text{km}\). In unknown-velocity experiments with Gaussian random fields, 8 events and 20 stations at \(CF \approx 0.94\) yield \(ACC=91.3\%\), \(e_\ell=0.07\,\text{s}\), velocity \(RMSE \approx 0.78\,\text{km/s}\), and location \(RMSE \approx 3.5\,\text{km}\). For 16 events, the paper reports \(ACC=72.0\%\) with 50 stations at \(CF \approx 0.98\) and \(ACC=28.4\%\) with 20 stations at \(CF \approx 0.90\). It also reports that increasing the number of stations improves velocity and location recovery at the cost of larger \(M\).

Robustness results show that when a distortion parameter \(p_d\) varies in
\[
c^* = (1-p_d)c_0 + p_d c_1,
\]
association accuracy remains above 90% for small confusion factor even as \(p_d \to 1\), and similar resilience is observed for interpolation between in-distribution and out-of-distribution velocity fields. With up to 30% removed picks and up to 30% random spurious picks, accuracy degradation is described as minor, approximately 5% at \(CF \approx 0.80\). The SGLD ablation is particularly sharp: in fixed-\(c\) experiments, SGD converges to local minima with \(ACC \approx 37.5\%\), whereas SGLD reaches \(ACC \approx 92.5\%\). The paper further states that HARPA is the first method to maintain greater than 90% ACC at \(CF>0.9\) and unknown three-dimensional \(c(x)\), a regime where PhaseLink, Bayesian GMMs, back-projection, and graph neural associations all fail.

The limitations are equally explicit. Current SGLD inference over \(O(10\text{–}50)\) events and \(O(20\text{–}50)\) stations takes minutes to hours on a GPU. The autoencoder spans a limited latent distribution, so larger-scale regional or global settings may require larger latent codes or physics-informed priors. The paper also identifies fully unbalanced OT in the global objective and integration with downstream tomography as open directions. Its stated implication is that HARPA enables routine microseismic monitoring in geothermal fields, hydraulic fracturing, and volcanoes, where dense clusters of tiny events carry structural and pre-failure information.

Source: https://www.emergentmind.com/topics/harpa