Modeling Authentic Research Trajectories: A Graph View of Research Ideation

Ruochen Li; Zhiyu Chen

cs.AIOct 19, 2026

Modeling Authentic Research Trajectories: A Graph View of Research Ideation

Authors: Ruochen Li, Zhiyu Chen

Organizations: Department of Computer Science, University of Texas at Dallas

Abstract

The research idea in a published paper is rarely the project's initial idea. Between the two lies a closed-loop process in which researchers run experiments, derive insights, identify challenges, explore the literature, and revise the idea, with each step motivated by evidence and reasoning they can articulate. We call the path one idea takes through this process, with the evidence and reasoning behind each step, its authentic research trajectory. Although it underlies every scientific contribution, it is rarely recorded. Existing LLM-based research agents generate an idea once, refine it against reviews, literature, or their own experimental results, or search over candidate ideas and experiments. In all of them the text of the idea evolves, while the evidence and reasoning that connected one version to the next are discarded. We propose the Research Progress Graph (RPG), a representation of authentic and complete research trajectories. An RPG records the evolution of an idea as typed research states connected by cognitive transitions, each annotated with a temporal index and the reasoning behind it, which makes this evolution an explicit object of study. We ground the representation in authentic trajectories whose steps their authors justify in annotations and papers. Comparing the representations of thirty-five research agents with the RPG, we find that each is an incomplete RPG; none keeps the state that changed, the evidence that changed it, and the reasoning connecting them as one evolving record. Modeling ideation with the RPG changes how research agents should be designed, evaluated, and trained. Agents predict the next research state instead of a complete idea, ideas are evaluated as trajectories instead of final texts, and authentic trajectories serve as training data. These directions shift research agents from generating static ideas to modeling how ideas evolve.

Figures & tables

Panel of Figure 1 — Figure 1: The Research Progress Graph (top) and the structures by which existing research agents represent ideation (bottom). An RPG records typed states, the temporal order of the transitions between them, and the reasoning process behind each transition, with results and literature entering as states that change the idea. Each existing structure keeps a subset of these elements and can be viewed as an incomplete RPG.

Type	Abbr.	Definition	Example
Research question	RQ	The central scientific problem that motivates the project	How to do table-to-text generation with only a few hundred examples?
Idea update	I	A new or revised research idea proposed at a particular stage of the project, often based on prior insights, challenges, literature, and results	A switch between a pre-trained LM and a copy mechanism, as in the pointer-generator network
Insight	IN	A key observation or higher-level conclusion derived from the current RPG	Copying is easy to learn from few shots; language modeling is not
Challenge	C	An open problem, bottleneck, or unanswered question that arises during the research process	How to combine a pre-trained LM with copying?
Literature search	LS	The act of searching prior work; holds the search query	“Few-shot language modeling”
Literature	L	A unit of literature, such as one paper, optionally with a summary, or an existing external resource, such as a repository of models and datasets	The pointer-generator paper

Table 1: State types of an RPG, with the running example of Case 1 (few-shot table-to-text generation).

Case	Idea type	Initial idea	Published idea	Rounds	Evidence
1	Method	Pointer-generator network for few-shot table-to-text	Pre-trained LM with a copy switch ( Chen et al., 2020 )	1	L (pointer-generator, re-read), R (own pilot), L (GPT-2), R (final)
2	Application	Prompt GPT-4 to role-play a patient	PATIENT- $\Psi$ , GPT-4 conditioned on clinician-written cognitive models and evaluated by mental-health experts and trainees ( Wang et al., 2024c ) ; then CBT-Bench on the therapist’s side ( Zhang et al., 2025 )	4	R (expert interviews), L (CBT), L (transcripts), R (user study), L (prior CBT tasks), R (benchmark)
3	Method (ideation agent)	Fine-tune an LLM on paper-to-idea pairs	LDC: reward models per dimension, dimensional controllers, sentence-level dynamic control ( Li et al., 2024a )	3	L (prompting-based ideation), L (ResearchAgent’s iterative feedback), L (trade-off in prior work), R $\times$ 2, R with human judges, R (later sweep)
4	System (research agent)	A prompted idea agent and an experiment agent that runs the source paper’s code	A ReAct experiment agent with retrieval and human feedback; the idea agent trained with fine-tuning and reward-model reinforcement learning ( Li et al., 2024b )	3	L (prior systems), L (MLAgentBench), L (ResearchAgent), R (agent drift), L (Hugging Face Hub), R (five tasks), R (single-prompt baseline), L (the group’s idea-generation method), R (retrained idea agent)
5	Empirical insight	Apply the one-shot RLVR lens to on-policy distillation	“Data-overfed but algorithm-starved”; inputs act as state generators ( Fu et al., 2026 )	3	L (one-shot RLVR, Part I), R $\times$ 3
6	Method	Fine-tune the student on teacher samples	On-policy distillation with per-token reverse KL, later used for recovery after mid-training (blog post, Lu and Thinking Machines Lab, 2025 )	3.5	R (SFT plateau), L (Qwen3 report), L (DAgger, GKD), R $\times$ 6

Table 2: Eight real ideas as RPGs. Cases 1 to 5 are presented in this section and Cases 6 to 8 in Appendix A . Rounds counts the passes through the cycle of § 2.1 ; Evidence lists the experimental result (R) and literature (L) states through which evidence entered.

Panel of Figure 2 — Figure 2: Case 1, few-shot table-to-text generation, as an RPG. One round, two experimental and two literature states of evidence. Transitions are numbered by temporal index; transition 1 has two sources, transition 10 re-reads a state already in the graph, and the reasoning process on transition 11 is shown. Listing in Appendix A.1 .

Panel of Figure 3 — Figure 3: Case 2, PATIENT- $\Psi$ and CBT-Bench, as an RPG. Four rounds, six states of evidence, two of them from people. Full account and listing in Appendix A.2 .

Panel of Figure 4 — Figure 4: Case 3, LDC, as an RPG. Three rounds, seven states of evidence, three of them from the literature. The trade-off enters from prior work at transition 4 and leads to both the controllers and the decoder (transition 6); the sweep that confirms it (Exp. Results #3) is a later analysis. Full account and listing in Appendix A.3 .

Panel of Figure 5 — Figure 5: Case 4, MLR-Copilot, as an RPG. Three rounds and an open challenge, nine states of evidence, five of them literature states. Literature #3 is an external resource, and Literature #4 is a later paper by the same group (Case 3). Full account and listing in Appendix A.4 .

Panel of Figure 6 — Figure 6: Case 5, on-policy distillation from one training example, as an RPG. Three rounds, four states of evidence. Full account and listing in Appendix A.5 .

Agent	v1	RQ	I	IN	C	LS	L	CE	R	Index	Reasoning
SciMON	2023-05	$\odot$	$\odot$	$\times$	$\times$	$\checkmark$	$\checkmark$	$\times$	$\times$	$\odot$	$\odot$
ResearchAgent	2024-04	$\odot$	$\checkmark$	$\odot$	$\times$	$\odot$	$\checkmark$	$\times$	$\times$	$\odot$	$\odot$
AI Scientist v1	2024-08	$\odot$	$\checkmark$	$\odot$	$\times$	$\odot$	$\odot$	$\checkmark$	$\checkmark$	$\odot$	$\times$
Si et al.	2024-09	$\odot$	$\checkmark$	$\times$	$\times$	$\odot$	$\odot$	$\times$	$\times$	$\times$	$\times$
Chain of Ideas	2024-10	$\odot$	$\checkmark$	$\odot$	$\times$	$\odot$	$\checkmark$	$\odot$	$\times$	$\odot$	$\odot$
Nova	2024-10	$\odot$	$\checkmark$	$\odot$	$\times$	$\checkmark$	$\checkmark$	$\times$	$\times$	$\odot$	$\odot$

Table 3: What twenty agents keep of an RPG, by element. v1 is the month of the first arXiv version, and the other columns are the eight state types of § 3.1 and the two edge annotations of § 3.2 : Index , the temporal index, read by the next step; Reasoning , the reasoning process bound to the change it explains.

\checkmark

a distinct object that a later step reads;

\odot

the information exists but is folded into a text, a prompt or a score, or is stored and not read;

\times

absent; ? not disclosed (Kosmos is closed source). Entries condense the group tables of § 5.2 to § 5.6 ; AIDE, R&D-Agent, and Gome appear in no group table, and their entries are read from their papers and code.

Agent	Idea representation	Loop: feedback $\rightarrow$ what is revised	P1	P2	P3	P4	P5
Si et al. ( Si et al., 2025b ; Si et al., 2025a )	Proposal JSON: title, problem, motivation, method, experiment plan, test cases, fallback	None; novelty filter only discards; closed-loop feedback named as future work	$\odot$	$\times$	$\times$	$\times$	$\times$
LDC ( Li et al., 2024a )	Text: problem, method, experiment plan, related work; input = a supporting paper	None; literature enters once before generation; novelty reward uses retrieval on the evaluation side only	$\odot$	$\times$	$\times$	$\times$	$\times$
Ramon Llull’s Thinking Machine ( Zhao et al., 2025b )	Theme $\times$ domain $\times$ method combination expanded to a title and abstract	None	$\odot$	$\times$	$\times$	$\times$	$\odot$
CycleResearcher ( Weng et al., 2025a )	Full paper text	Training-time only: review scores train the next generator	$\times$	$\times$	$\times$	$\times$	$\times$
ResearchAgent ( Baek et al., 2025 )	Problem, method, experiment, each with rationale	Yes: 15 reviewing agents, 5 criteria $\rightarrow$ whole-section regeneration, 3 rounds to saturation; no evidence enters the loop	$\odot$	$\odot$	$\odot$	$\times$	$\times$
ResearchTown ( Yu et al., 2025a )	Typed objects: insight, idea, proposal, review, and meta-review, with keys and timesteps; a rebuttal type is defined but its generation is disabled in the code	None at run time; the review $\rightarrow$ proposal transition is declared and never called	$\odot$	$\odot$	$\odot$	$\times$	$\times$

Table 4: Group A: a single idea text, generated once or revised against the model’s own critique. Scores:

\checkmark

full,

\odot

partial,

\times

absent; P1 to P5 are the properties of § 3.4 ; in column P4, lit, exp, and both name the evidence that enters after the initial idea and changes it, and

\odot

marks evidence that changes only the code, the plan, or a hypothesis inside a fixed question; ? marks a score that cannot be given because the representation is not disclosed.

Agent	Idea representation	Loop	P1	P2	P3	P4	P5
SciMON ( Wang et al., 2024b )	One sentence with a record of input, prediction, nearest prior work, regeneration	Literature similarity $\rightarrow$ regeneration until no retrieved work is too close	$\odot$	$\odot$	$\odot$	lit	$\times$
MOOSE-Star ( Yang and Bing, 2026 )	Hypothesis built in increments from a research background, one increment (motivation, mechanism, methodology) per inspiration paper	Motivation $\rightarrow$ hierarchical search for the next inspiration $\rightarrow$ increment composed on the previous hypothesis, one step per inspiration; models trained on 108,717 papers decomposed by an LLM; no feedback loop	$\odot$	$\odot$	$\odot$	lit	$\times$
Chain of Ideas ( Li et al., 2025b )	Chain of papers in development order; idea = motivation, novelty, method	Novelty check $\rightarrow$ refine from chain; review + retrieval $\rightarrow$ experiment plan, 1 round	$\odot$	$\odot$	$\odot$	lit	$\times$
Nova ( Hu et al., 2024 )	Idea JSON with parent pointer	Search plan on current idea $\rightarrow$ child ideas, 3 rounds; tree stored, not read back	$\odot$	$\odot$	$\odot$	lit	$\times$
VirSci ( Su et al., 2025 )	Idea JSON with title, experiment, and self-ratings $\rightarrow$ abstract	Team discussion $\rightarrow$ idea; similarity $\rightarrow$ abstract	$\odot$	$\odot$	$\odot$	lit	$\times$
AI Scientist v2 ( Yamada et al., 2025 )	Idea JSON frozen to a file; experiment tree nodes typed	Literature $\rightarrow$ proposal in ideation only; results $\rightarrow$ experiment tree, never the idea	$\odot$	$\odot$	$\odot$	lit	$\odot$

Table 5: Group B: an idea text with literature retrieved for it. Results, where experiments are run, do not revise the idea.

Agent	Idea representation	Loop	P1	P2	P3	P4	P5
AI Scientist v1 ( Lu et al., 2024 )	Idea JSON	Up to two self-reflection rounds per idea in the released pipeline; results $\rightarrow$ next experiment code and paper; score feedback to ideas described but not wired in the released code	$\odot$	$\odot$	$\times$	$\odot$	$\times$
AI-Researcher ( Tang et al., 2025 )	Sectioned proposal	Advisor and judge $\rightarrow$ code	$\odot$	$\odot$	$\odot$	$\odot$	$\times$
Agent Laboratory ( Schmidgall et al., 2025 )	Text topic and plan	LLM review + last results $\rightarrow$ new plan; human checkpoints	$\odot$	$\odot$	$\odot$	$\odot$	$\checkmark$
Curie ( Kon et al., 2025 )	Experiment knowledge module with a DAG	Results $\rightarrow$ hypothesis within a fixed experimental question	$\odot$	$\checkmark$	$\odot$	$\odot$	$\times$
Jr. AI Scientist ( Miyai et al., 2025 )	Idea JSON as AI Scientist v2	Results $\rightarrow$ code; hypothesis fixed	$\odot$	$\odot$	$\times$	$\odot$	$\times$
MLR-Copilot ( Li et al., 2024b )	Task, related work, hypothesis, plan	Results $\rightarrow$ researcher $\rightarrow$ hypothesis	$\odot$	$\odot$	$\times$	$\odot$	$\checkmark$

Table 6: Group C: a fixed idea with a tree of experiments; results change the code, the plan, or a hypothesis inside a fixed question, not the idea.

Agent	Idea representation	Loop	P1	P2	P3	P4	P5
Dolphin ( Yuan et al., 2025 )	Idea JSON	Result label (improvement, maintenance, decline) $\rightarrow$ improved ideas marked effective in the next prompt, others used as a similarity filter; new ideas, not revisions; numbers and analysis do not reach the prompt	$\odot$	$\odot$	$\times$	exp	$\times$
DeepScientist ( Weng et al., 2025b )	Findings memory typed by lifecycle stage	Results + analysis records retrieved by the next hypothesis step	$\odot$	$\checkmark$	$\odot$	both	$\odot$
Kosmos ( Mitchener et al., 2025 )	Structured world model, format not public	Literature + data analysis summaries $\rightarrow$ world model $\rightarrow$ next tasks	?	$\odot$	?	both	$\times$
ASI-Arch ( Liu et al., 2025 )	Node: motivation, code, results, analysis, retrieved cognition; parent pointer	Results + shortcoming analysis + retrieved cognition $\rightarrow$ child architecture	$\odot$	$\checkmark$	$\odot$	both	$\times$
AlphaResearch ( Yu et al., 2025b )	Ordered trajectory of idea, program, and result entries	Results $\rightarrow$ next idea; a simulated review score rejects ideas before they run	$\odot$	$\odot$	$\times$	exp	$\times$
freephdlabor ( Li et al., 2025a )	Working idea JSON	Free-text results $\rightarrow$ manager $\rightarrow$ re-ideation; human can interject	$\odot$	$\odot$	$\times$	exp	$\checkmark$

Table 7: Group D: ideas revised from experimental results.

Agent	Idea representation	Loop	P1	P2	P3	P4	P5
Evolving Idea Graphs ( Dong et al., 2026 )	Typed nodes (problem, hypothesis, method, assumption, risk, evaluation plan, novelty claim, evidence need, repair) and typed edges; every operation logged with rationale and timestamp	Controller edit and commit steps inside the graph; literature given in advance; no retrieval, no experiments	$\checkmark$	$\odot$	$\checkmark$	$\times$	$\times$

Table 8: Group E: typed idea graphs that no evidence enters.

Criterion	Authors’ version	One-shot rewrite	Inconsistent	$p$ (rewrite) [95% CI]
Novelty	19	3	4	0.14 [0.05, 0.33]
Excitement	21	4	1	0.16 [0.06, 0.35]
Feasibility	8	16	2	0.67 [0.47, 0.82]
Effectiveness	10	13	3	0.57 [0.37, 0.74]
Overall	13	10	3	0.43 [0.26, 0.63]
Novelty, verified	13	9	4	0.41 [0.23, 0.61]

Table 9: Blind judges on 26 pairs, each pairing the authors’ own next version with a tool-free one-shot rewrite of the same seed. The first five rows are the text-only judge; the last row (“Novelty, verified”) is the verified-novelty judge, whose retrieval is restricted to work posted before the target. A preference counts only when both presentation orders agree; the remainder are inconsistent.

p

(rewrite) is the share of decisive pairs in which the judge preferred the rewrite, with a 95% Wilson interval.

Recommendation	Problem addressed	Shown in	What is gained
Store the trajectory	Histories are written and not read; the loop passes a score, a label, a review, or a summary, and regenerates the whole idea	§ 5.7 ; Case 1	Every change traced to a component, the evidence that caused it, and a reason; failed ideas kept with the reason they failed
Evaluate relatively, with a grounded judge	A text-only judge rewards unchecked claims and penalizes cited ones	Case 2; pilot and blind study (§ 6.4 )	Grounded ideas are no longer penalized; the effect of the representation is separated from the effect of compute
Train on next-state prediction	Whole-idea training gives one signal per idea, and that signal comes from a biased judge	§ 6.1 , § 4.6 ; MOOSE-Star	Supervision per step that a person can audit; no reward model over whole ideas
Annotate update paths	Training corpora hold final ideas only	§ 2.1 ; the eight cases	Training data for next-state prediction and reference trajectories for evaluation
Record agents and people on one graph	Human input and experimental evidence are recorded apart from each other and from the idea	§ 5.7	One record of the project; the human role defined per transition

Table 10: The five recommendations of this section, the problem each addresses, where this paper shows the problem, and what is gained by following it.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

id	type	content (short)
RQ0	RQ	How to do table-to-text NLG with only a few hundred examples
L0	L	Pointer-generator network (See et al. 2017), built on pointer networks; template NLG baselines
I0	I	Pointer-generator network trained on the few-shot set
CE0	CE	Implement pointer-generator; train at 50/100/200/500 examples
R0	R	Copies table values well; connective language is poor at all sizes
IN0	IN	Two skills are needed: copy (easy from few shots) and language modeling (hard from few shots)

Appendix

Table 17

#	edge	transition
1	RQ0, L0 $\rightarrow$ I0	Choose the standard architecture for copy-heavy generation
2	I0 $\rightarrow$ CE0	Implement the pointer-generator
3	CE0 $\rightarrow$ R0	Train and evaluate at 50 to 500 examples
4	R0 $\rightarrow$ IN0	Analyze errors; copying succeeded, fluency failed
5	IN0 $\rightarrow$ C1	Isolate the hard skill as the challenge
6	C1 $\rightarrow$ LS1	Search for ways to get language modeling cheaply

Appendix

Table 18

id	type	content (short)	source
RQ0	RQ	How can CBT trainees practice formulating a patient’s cognitive model before seeing real patients	Wang et al. 2024c
I0	I	Prompt GPT-4 to role-play a patient with a diagnosis	Wang et al. 2024c
R0	R (formative interviews, 12 experts and trainees)	Training did not prepare them for real practice; peer role-play is unrealistic; real patients show different conversational styles	Wang et al. 2024c §1, App. B
IN0	IN	Realism = behavior generated from a structured cognition; trainees must infer beliefs from behavior	Wang et al. 2024c
C1	C	How to give an LLM a structured cognition	Wang et al. 2024c
LS1	LS	Search in clinical psychology: how CBT models a patient	Wang et al. 2024c

Appendix

Table 19

#	edge	transition
1	RQ0 $\rightarrow$ I0	Default choice
2	I0 $\rightarrow$ R0	Formative interviews with experts and trainees instead of a run
3	R0 $\rightarrow$ IN0 $\rightarrow$ C1	Diagnose lack of realism as lack of structured cognition (reason annotated, or traced and confirmed, by the authors; the paper states that a cognitive model gives high fidelity, §1)
4	C1 $\rightarrow$ LS1 $\rightarrow$ L1 $\rightarrow$ IN1	Search a different field; identify the CCD as the parameterization an LLM can simulate
5	IN1, R0 $\rightarrow$ I1	Combine CCD conditioning with the styles experts named
6	I1 $\rightarrow$ C2 $\rightarrow$ L2 $\rightarrow$ CE1	Build the missing data with clinicians

Appendix

Table 20

id	type	content (short)	source
RQ0	RQ	How to generate research ideas that meet expert standards of novelty, feasibility and effectiveness	Li et al. §1
L0	L	Ideation systems built on pre-trained models without task-specific learning (e.g., ResearchAgent)	Li et al. §1, §2
L2	L	ResearchAgent: ideas refined iteratively with feedback from reviewing agents on several criteria	annotated, or traced and confirmed, by the authors
IN0	IN	Prompting alone leaves outputs “misaligned with expert standards”; the task must be learned	Li et al. abstract (quoted), §1
I0	I	Supervised fine-tuning on 1,000 (paper $\rightarrow$ extracted idea) pairs; supporting work chosen by prompting with citation counts	Li et al. §3.2, §4.1
IN1	IN	SFT gives the task format and RL stability; the quality dimensions are left to RL	Li et al. §1, §3.2 (stated purpose of the SFT stage)

Appendix

Table 21

#	edge	transition
1	RQ0, L0 $\rightarrow$ IN0 $\rightarrow$ I0	Read prompting-based ideation as unlearned; fine-tune
2	I0 $\rightarrow$ IN1 $\rightarrow$ C1	State what SFT does and does not optimize; pose the multi-dimension challenge
3	C1, L2 $\rightarrow$ I1 $\rightarrow$ R0	Add per-dimension reward models, with the feedback design taken from ResearchAgent (annotated, or traced and confirmed, by the authors); measure
4	L1 $\rightarrow$ IN2 $\rightarrow$ C2	Read the trade-off reported in prior work as a property of the task; ask for separate control (reason stated in §1 and §3.4)
5	C2 $\rightarrow$ I2 $\rightarrow$ R1	Add dimensional controllers; measure
6	IN2, R1 $\rightarrow$ IN3 $\rightarrow$ I3	A fixed setting trades one dimension against another across the whole idea, so localize emphasis to parts of the idea; add the sentence-level decoder (trade-off annotated, or traced and confirmed, by the authors; parts stated in §1 and §5.2)

Appendix

Table 22

id	type	content (short)	source
RQ0	RQ	Carry ML research from an existing paper to executed experiments automatically	v1 §1
L0	L	Ideation systems whose ideas are not executed; experiment automation limited to small edits, without robust feedback	v1 §1; v3 §1
C1	C	Connect ideas to code that runs	v1 §1, §2
I0	I	Prompted idea agent (paper + related work $\rightarrow$ hypothesis $\rightarrow$ plan); experiment agent adapts and runs the paper’s prototype code; no model or data retrieval, no human feedback	annotated, or traced and confirmed, by the authors
L1	L	MLAgentBench: a research agent built on the ReAct framework (thought, action, observation)	annotated, or traced and confirmed, by the authors
L2	L	ResearchAgent: iterative feedback that refines the work	annotated, or traced and confirmed, by the authors

Appendix

Table 23

#	edge	transition
1	RQ0, L0 $\rightarrow$ C1	Prior systems either do not execute ideas or only make small edits (annotated, or traced and confirmed, by the authors)
2	C1 $\rightarrow$ I0	Start execution from the paper’s own prototype code (annotated, or traced and confirmed, by the authors)
3	I0, L1, L2 $\rightarrow$ I1	Take the ReAct framework from MLAgentBench’s agent and the iterative feedback from ResearchAgent (annotated, or traced and confirmed, by the authors)
4	I1 $\rightarrow$ R0 $\rightarrow$ IN1	The agent drifts; a person can correct the direction (annotated, or traced and confirmed, by the authors)
5	L3, IN1 $\rightarrow$ I2	Add retrieval from the Hub, on one researcher’s idea that it matches models to the task better, and human feedback to correct the drift (annotated, or traced and confirmed, by the authors)
6	I2 $\rightarrow$ R1	Evaluate on five tasks

Appendix

Table 24

id	type	content (short)
RQ0	RQ	How much training data does on-policy distillation need, and why
L0	L	One-shot RLVR as a controlled experiment isolating the role of data ( Wang et al., 2025b ) ; Part I ( Li et al., 2026 ) , cited in Part II for OPD’s training dynamics and the top- $k$ overlap metric
I0	I	Bring the one-shot lens to OPD: train on a single query (no mechanism, no prediction, no boundary yet)
CE0	CE	One-shot OPD on one query per domain for hundreds of steps; full-data OPD as reference (17k queries in math); 4 domains (math, code, instruction following, agentic tool use) $\times$ 3 families (Qwen, Llama, OLMo)
R0	R	One-shot keeps improving for hundreds of steps; 68.5 vs 69.8 at step 300 (three-benchmark average); 69% of the teacher–student gap and 87% of full-data gain recovered
IN0	IN	State coverage (fraction of 200 state clusters reached): one query’s rollouts reach 65.9% by step 100 and 71.5% by step 300; supervision breadth comes from rollouts, not from prompts

Appendix

Table 25

#	edge	transition
1	L0 $\rightarrow$ I0	Transfer an experimental lens from RLVR to OPD
2	I0 $\rightarrow$ CE0 $\rightarrow$ R0	Sharpen “how much data” into a measurable one-shot experiment
3	R0 $\rightarrow$ IN0	Mechanize: explain the gain by state coverage
4	IN0 $\rightarrow$ C1	Derive a discriminating question from the mechanism
5	C1 $\rightarrow$ CE1 $\rightarrow$ R1 $\rightarrow$ IN1	Ablate content vs states; confirm inputs are state generators
6	IN1 $\rightarrow$ C2	The mechanism leaves a residual: slow absorption

Appendix

Table 26

Panel of Figure 7 — Figure 7: Case 6, on-policy distillation, as an RPG. Three and a half rounds, seven experimental and two literature states of evidence, three idea updates after the initial one. Listing below.

id	type	content (short)	section of the post
RQ0	RQ	Give a small model the reasoning ability of a large one at low cost	Distillation for reasoning
L0	L	Qwen3 report: on-policy distillation reaches 74.4% AIME’24 in 1,800 GPU hours vs 17,920 for RL	Distillation for reasoning: Reinforcement learning
I0	I	Off-policy distillation: sample from the teacher, SFT the student	Distillation for reasoning: Off-policy distillation
CE0	CE	SFT on 400k teacher prompts	Off-policy distillation
R0	R	60% AIME’24 at 400k; extrapolated ~2M prompts for 70%	Off-policy distillation
L1	L	DAgger; GKD; MiniLLM; process reward models	On-policy distillation — best of both worlds

Appendix

Table 28

#	edge	transition
1	RQ0 $\rightarrow$ I0 $\rightarrow$ CE0 $\rightarrow$ R0	Run the default and extrapolate its cost
2	R0, L0, L1 $\rightarrow$ IN0	Diagnose both baselines; off-policy compounds error, RL is sparse (the post states the diagnosis in its introduction)
3	IN0 $\rightarrow$ I1	Combine on-policy sampling with dense teacher grading
4	I1 $\rightarrow$ CE1 $\rightarrow$ R1	Run from the same checkpoint; measure cost
5	R1 $\rightarrow$ C1 $\rightarrow$ CE2 $\rightarrow$ R2	Isolate dense supervision by distilling an RL policy back into its base
6	R2 $\rightarrow$ IN1	Reframe RL as search

Appendix

Table 29

Panel of Figure 8 — Figure 8: Case 7, from the ideation study to the ideation–execution gap, as an RPG. One round and an open challenge, five states of evidence. Listing below.

id	type	content (short)	source
RQ0	RQ	Can LLMs generate research ideas as good as experts’?	Si et al. 2025 §1
L0	L	Ideation systems and their evaluations: small studies or LLM judges	Si et al. 2025 §1, §2
I0	I	A large blind human study; LLM agent = RAG + 4,000 seeds per topic + dedup + Swiss-tournament ranker	Si et al. 2025 §3
CE0	CE	100+ NLP researchers write and review; 79 reviewers; blind	Si et al. 2025 §4
R0	R	AI ideas more novel (5.64 vs 4.84); feasibility slightly lower, not significant	Si et al. 2025 abstract, §5.1
R1	R (self-analysis)	about 200 unique of 4,000 ideas; ranker 53.3% agreement with humans; reviewer–reviewer agreement 56.1%	Si et al. 2025 §7.1, §7.2, §6.3

Appendix

Table 31

#	edge	transition
1	RQ0, L0 $\rightarrow$ I0 $\rightarrow$ CE0 $\rightarrow$ R0	Replace small or automatic evaluation by a large blind human study; measure
2	R0 $\rightarrow$ R1	Analyze the study’s own generator and judges
3	R1 $\rightarrow$ IN0 $\rightarrow$ C1	Read the agreement and diversity numbers as a limit of idea-level judgment; ask the execution question
4	C1 $\rightarrow$ I1 $\rightarrow$ CE1 $\rightarrow$ R2	Execute a subset of the same ideas; review blind; measure the drop
5	R2 $\rightarrow$ R3 $\rightarrow$ IN1	Rule out executor changes and evaluation substitution; locate the gap in the ideas and the judge
6	IN1 $\rightarrow$ C2	Leave the loop open: outcomes should feed back into idea improvement

Appendix

Table 32

Panel of Figure 9 — Figure 9: Case 8, DeepSeek-R1, as an RPG. Abandoned updates are drawn as dashed transitions. Three rounds, two recorded abandoned branches, one fork, five states of evidence. The report places the abandoned attempts only in the early stages of the project. Listing below.

id	type	content (short)	source
RQ0	RQ	How to improve LLM reasoning with reinforcement learning at scale	R1 §1
L0	L	DeepSeekMath: GRPO (no critic, group-relative baseline) after SFT on mathematics; the group’s own prior work	DeepSeekMath §4; R1 §2.2.1
IN0	IN	Earlier RL for reasoning “heavily depended on supervised data”; explore reasoning “without any supervised data”, through pure RL	R1 §2.2 (quoted)
C1	C	Can reasoning emerge from RL on the base model with no supervised data? How to supervise it?	R1 §2.2, §2.2.2
I1	I	DeepSeek-R1-Zero: GRPO on DeepSeek-V3-Base; rule-based accuracy + format rewards; think/answer template; no neural reward model (reward hacking)	R1 §2.2.1–2.2.3
R1	R	AIME 2024 pass@1 15.6% $\rightarrow$ 71.0%, 86.7% majority voting; self-evolution: longer thinking, reflection, “aha moment”	R1 §2.2.4

Appendix

Table 34

#	edge	transition
1	RQ0, L0 $\rightarrow$ IN0 $\rightarrow$ C1	State the aim of developing reasoning without supervised data; ask whether RL alone suffices and how to supervise it
2	C1 $\rightarrow$ I1 $\rightarrow$ R1, R2	Run pure RL with rule-based rewards; observe the capability and the drawbacks
2a	C1 $\dashrightarrow$ X1	Try a process reward model; abandon, reasons recorded
2b	C1 $\dashrightarrow$ X2	Try MCTS; abandon, reasons recorded
3	R1, R2 $\rightarrow$ IN1 $\rightarrow$ C2	Read the promising results and the drawbacks together; expect cold-start data to fix readability and raise performance
4	C2 $\rightarrow$ I2 $\rightarrow$ R3	Add cold start and three more stages; pay a small accuracy cost for readability

Appendix

Table 35

Case	Type	Who states the reasons	Rounds	Evidence from experiments or studies	Evidence from literature
1	method	the authors (annotated, or traced and confirmed)	1	2	2
2	application	the authors (annotated, or traced and confirmed)	4	3 (formative interviews, user study, benchmark)	3 (one is a data source)
3	method (ideation system)	the paper and the authors (annotated, or traced and confirmed), in the paper’s order of presentation	3	4	3
4	system (research agent)	the paper, two versions, and the authors (annotated, or traced and confirmed)	3 + an open challenge	4 (one an observation of the agent drifting)	5 (one an external resource, one the group’s own later paper)
5	empirical insight	the papers (shared title, eight shared authors)	3	3	1
6	method	the authors (blog)	3.5	7	2

Appendix

Table 36

Case	Decisive cognitive pattern	Node changed / evidence / recorded reason	What a loop that regenerates the whole idea from a score or summary keeps
1	abstract structure, substitute one slot	LM slot of the switch / R0 + re-read L0 / pointer-generator is a switch between LM and copy	the final model, not the attribution
2	transfer a parameterized cognitive model; build the missing data; move from patient to therapist	conditioning input, styles, then the object (patient $\rightarrow$ therapist) / R0 + L1, R0, R1 + IN2 / structured cognition; expert-named styles; the cognitive model is central on both sides	a better prompt, not the conclusion about the judge or the move to the therapist
3	take a trade-off from prior work as a property of the task; give each dimension its own handle; localize	the generator (per-dimension rewards $\rightarrow$ per-dimension controllers), then where emphasis applies / L1 (trade-off in prior work), for both / the dimensions are interdependent; parts of an idea need different emphasis	a better reward model, not the prior-work trade-off that motivated the controllers and the decoder
4	take components from other papers one at a time; add a person where the agent drifts; adopt a better method once it exists	the experiment agent, then its resources and a human channel, then the idea agent (prompted $\rightarrow$ trained) / L1 + L2, then L3 + R0, then L4 / ReAct from MLAgentBench and feedback from ResearchAgent; retrieval from the Hub matches models better, and a person corrects the drift; the better ideation method now exists	a better system, not which component came from which paper or why each was added
5	sharpen, mechanize, ablate what the mechanism predicts	claim scope, twice / R1, R2 / IN0 predicts content-independence; residual slow absorption	a good score, not the two ablations the mechanism predicted
6	diagnose both baselines, combine; then rule out alternatives by experiment	sampling source and grading, then the training target / R0 + L0 + L1, then R3 + R4 / compounding error and sparse reward; finite-batch drift	a data-mixing recipe, not the experiment that ruled mixing out

Appendix

Table 37

Condition	Loop type	Procedure	Corresponds to
One-shot rewrite	none	One rewrite with the whole budget	generate-score-rewrite systems
Review loop	reviewer feedback	Rounds of critique and whole-idea rewrite until the budget is spent, no tools	ResearchAgent-style
Literature loop	retrieval	as above, with one search per round, up to Q, whose results are visible next round; whole-idea rewrite	SciMON / Nova-style
RPG loop	typed graph	Transitions, until the budget is spent, of type insight, challenge, search or idea update; the search is executed by the same tool and its results become literature nodes; an idea update is a localized edit of named schema fields, each citing an insight or literature node created since the previous update, with untouched fields kept verbatim; an edit log records which field changed because of which node	this paper
RPG without outside information	typed graph, closed	as above with search and literature removed	ablation

Appendix

Table 11: The five equal-budget conditions.

Condition	Calls	Out. tok.	In. tok.	Wins vs RPG
One-shot rewrite	1.0	2,096	5,274	1 of 3
Review loop	5.7	14,026	37,997	3 of 3
Literature loop	6.0	13,430	43,828	3 of 3
RPG loop	6.0	5,006	51,031	—
RPG, no outside info.	6.0	5,793	45,363	0 of 3

Appendix

Table 12: Pilot (n = 3 seeds): mean spend per run and overall wins against the RPG loop.

Research Idea Generation

Research idea generation uses computational methods, especially language models, to propose and assess new questions, hypotheses, and research directions.

See all 60 papers

Jul 11, 2026

IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

Scientific ideation unfolds over multiple stages, including literature search, paper reading, tool use, claim checking, cross-paper synthesis, brainstorming, rejection of weak directions, and iterative writing. Yet most existing resources capture isolated components or final artifacts rather than the process connecting them. We introduce IdeaTrail, a dataset of 1,170 multi-turn trajectories for scientific ideation and proposal generation. Each trajectory follows a research process from evidence gathering to either idea selection or proposal construction, jointly recording tool use, acquired evidence, intermediate artifacts, and reasoning. IdeaTrail is synthesized from human-selected research papers and proposal artifacts through a Generator--Advisor loop. The Generator produces the visible sequence of actions, observations, and artifact edits, while the Advisor uses the full generation context to check grounding, causal order, naturalness, and leakage from hidden targets. This reverse-to-forward design keeps trajectories aligned with real scientific artifacts while retaining the uncertainty, evidence use, and staged convergence characteristic of research practice. IdeaTrail provides both reusable process supervision and a general recipe for constructing scientific-research-agent data.

Hengquan Guo

ShanghaiTech University

AI Agents for Scientific Discovery Scientific Hypothesis Generation LLM Agent Trajectory Synthesis Research Idea Generation Synthetic Data Generation for Language Models

Ask PDF

Aug 11, 2026

Tree-of-Ideas: Automated Research Ideation via Cross-Trajectory Reasoning over Scholarly Evolution

Effective research ideation requires moving beyond a static understanding of prior work to trace how research problems and solutions evolve across the literature. Existing methods either treat papers as unstructured context or model scholarly evolution as isolated citation chains, overlooking interactions among research trajectories. We propose Tree-of-Ideas (ToI), a two-stage framework. EvoTrace reconstructs branching scholarly trajectories from citations, tracking evolving methods, resolved problems, and gaps. EvoAgent then reasons across trajectories to identify convergent problems and complementary solutions, generating grounded research ideas. Across six AI research topics, ToI achieves the highest score among automatic methods (6.27 vs. 5.36 for the strongest baseline on a 10-point scale), with strong Novelty (6.36) and Groundedness (7.00). Also, its score approaches that of human-paper references (6.29), demonstrating the value of cross-path evolutionary reasoning.

Xun Li, Yiying Yang, Pengtao Li +8

Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · University of Melbourne · University of Sydney +4

AI Agents for Scientific Discovery Research Idea Generation Large Language Model-Guided Scientific Discovery

Ask PDF

May 1, 2026

Structure Liberates: How Constrained Sensemaking Produces More Novel Research Output

Scientific discovery is an extended process of ideation -- surveying prior work, forming hypotheses, and refining reasoning -- yet existing LLM research agents treat this phase as a brief preamble despite its central role in research. We introduce SciSense, a sensemaking-grounded framework that operationalizes ideation as a structured sequence of eight cognitive stages (Pirolli and Card, 2005). We construct SciSense-Traj, a dataset of 120K citation-conditioned research trajectories in two modes: Hindsight, where an LLM reconstructs the ideation path from a paper's cited works to the paper itself, and Foresight, where the LLM proposes novel directions from the same citations. We distill these into SciSense-LM, a family of sensemaking LLMs spanning 3B to 70B parameters. Contrary to the assumption that looser supervision promotes exploration, Hindsight-trained models improve trajectory quality by 2.0% over Foresight-trained models while also producing more novel and diverse outputs, across all eight model families we test. The advantage propagates downstream: coding agents conditioned on Hindsight trajectories produce research artifacts with higher executability, scientific grounding, and utility than those conditioned on Foresight trajectories. This suggests that outcome-anchored ideation reduces the cognitive burden on downstream agents, freeing them to explore more creatively. SciSense offers both a practical tool for augmenting LLM-driven research workflows and a principled testbed for studying how planning shapes scientific discovery.

James Mooney, Zae Myung Kim, Young-Jun Lee +1

University of Minnesota

AI Agents for Scientific Discovery Scientific Hypothesis Generation Research Idea Generation Large Language Model-Guided Scientific Discovery

Ask PDF

Abstract

Research Idea Generation

Explore similar work

IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

Tree-of-Ideas: Automated Research Ideation via Cross-Trajectory Reasoning over Scholarly Evolution

Structure Liberates: How Constrained Sensemaking Produces More Novel Research Output