literature-review-agent

Name: literature-review-agent
Author: woodfishhhh

31stars1forks

How to install literature-review-agent?

woodfishhhh/ez_math_model/literature-review-agent

$npx -y skills add woodfishhhh/EZ_math_model --skill literature-review-agent

Installs into the current project.

Summary

Step 3 of the PaperOrchestra pipeline (arXiv:2604.05018). Execute the literature search strategy from outline.json — discover candidate papers via web search, verify them through Semantic Scholar (Levenshtein > 70 fuzzy title match, temporal cutoff, dedup by paperId), build a Bib

Files · 1

View on GitHub

SKILL.md

1	# Literature Review Agent (Step 3)
2
3	Faithful implementation of the Hybrid Literature Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 3, App. D.3, App. F.1 p.46).
5
6	Cost: ~20–30 LLM calls. This is one of the two longest steps (the other is
7	plotting). Wall-time floor is set by Semantic Scholar's 1 QPS verification
8	limit.
9
10	## Inputs
11
12	- `workspace/outline.json` — specifically `intro_related_work_plan` with the
13	Introduction search directions and the 2-4 Related Work methodology
14	clusters
15	- `workspace/inputs/conference_guidelines.md` — used to derive `cutoff_date`
16	- `workspace/inputs/idea.md`, `workspace/inputs/experimental_log.md` — for
17	framing the Intro and grounding the Related Work positioning
18
19	## Outputs
20
21	- `workspace/citation_pool.json` — verified Semantic Scholar metadata for
22	every paper that survived verification
23	- `workspace/refs.bib` — BibTeX file generated from the verified pool
24	- `workspace/drafts/intro_relwork.tex` — drafted Introduction and Related
25	Work sections, written into the template, with the rest of the template
26	preserved verbatim
27
28	## Two-phase pipeline (App. D.3)
29
30	```
31	PHASE 1 — Parallel Candidate Discovery
32	For each search direction in introduction_strategy.search_directions:
33	For each limitation_search_query in each related_work cluster:
34	- Use the host's web search tool to discover up to ~10 candidate papers.
35	- Run up to 10 discovery queries in parallel (host-permitting).
36	- Collect (title, snippet, url) tuples — no verification yet.
37	→ PRE-DEDUP before Phase 2 (see Step 1.5 below)
38
39	PHASE 2 — Sequential Citation Verification (1 QPS, with cache)
40	For each candidate (after pre-dedup), sequentially:
41	0. Check s2_cache.json first (scripts/s2_cache.py --check).
42	If HIT: use cached response, skip live S2 call. No throttle needed.
43	If MISS: proceed with live request below.
44	1. Query Semantic Scholar by title:
45	GET https://api.semanticscholar.org/graph/v1/paper/search?query=<title>
46	&fields=title,abstract,year,authors,venue,externalIds&limit=5
47	(Public endpoint, no key. Throttle to 1 QPS for live requests only.)
48	2. Store the S2 response in cache: s2_cache.py --store.
49	3. Pick the top hit. Check Levenshtein title ratio against the original
50	candidate title. If ratio < 70: discard.
51	4. Bonus: if year and venue exactly align with hints, add a +5 point
52	match-quality bonus.
53	5. Require: abstract is non-empty.
54	6. Require: paper.year (or month if known) strictly predates cutoff_date.
55	Months default to day-1: e.g., "October 2024" → 2024-10-01.
56	7. If all checks pass, add to verified pool.
57	After all candidates are verified, dedup by Semantic Scholar paperId.
58	```
59
60	The host agent does the LLM/web work; the deterministic helpers in `scripts/`
61	do the math.
62
63	## Step-by-step
64
65	### 0. Derive `cutoff_date`
66
67	Parse `conference_guidelines.md` for the submission deadline. The paper aligns
68	research cutoff with venue submission deadline (App. D.1):
69
70	\| Venue \| Cutoff \|
71	\|---\|---\|
72	\| CVPR 2025 \| Nov 2024 \|
73	\| ICLR 2025 \| Oct 2024 \|
74	\| Other \| One month before the stated submission deadline \|
75
76	Encode as `YYYY-MM-DD`. Months default to day-1 (e.g., `2024-10-01`).
77
78	### 1. Phase 1: Parallel Candidate Discovery
79
80	From `outline.json`:
81
82	- All `introduction_strategy.search_directions` (3-5 queries)
83	- For each cluster in `related_work_strategy.subsections`:
84	- The cluster's `sota_investigation_mission` becomes a search query
85	- All `limitation_search_queries` (1-3 each)
86
87	For each query, use your host's web search tool (e.g., `WebSearch` in
88	Claude Code, `@web` in Cursor, the search tool in Antigravity). Collect the
89	top ~10 candidates per query: title, abstract snippet, source URL.
90
91	If your host supports parallel sub-tasks, fire up to 10 concurrent search
92	queries. If not, run sequentially — slower but functionally equivalent.
93
94	#### Optional: Exa as a Phase 1 backend
95
96	If your host has no native web search, OR you want a research-paper-focused
97	backend with better signal-to-noise, you can use [Exa](https://exa.ai) via
98	the bundled `scripts/exa_search.py` helper. It is opt-in and reads
99	`EXA_API_KEY` from the environment — the repo never commits a key.
100
101	```bash
102	export EXA_API_KEY="your-key-here" # get one at https://dashboard.exa.ai/
103	python skills/literature-review-agent/s

People who liked this also viewed

6 picks

Type

literature-review-agent

31stars1forks

View on GitHub

How to install literature-review-agent?

woodfishhhh/ez_math_model/literature-review-agent

$npx -y skills add woodfishhhh/EZ_math_model --skill literature-review-agent

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Literature Review Agent (Step 3)
2
3	Faithful implementation of the Hybrid Literature Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 3, App. D.3, App. F.1 p.46).
5
6	Cost: ~20–30 LLM calls. This is one of the two longest steps (the other is
7	plotting). Wall-time floor is set by Semantic Scholar's 1 QPS verification
8	limit.
9
10	## Inputs
11
12	- `workspace/outline.json` — specifically `intro_related_work_plan` with the
13	Introduction search directions and the 2-4 Related Work methodology
14	clusters
15	- `workspace/inputs/conference_guidelines.md` — used to derive `cutoff_date`
16	- `workspace/inputs/idea.md`, `workspace/inputs/experimental_log.md` — for
17	framing the Intro and grounding the Related Work positioning
18
19	## Outputs
20
21	- `workspace/citation_pool.json` — verified Semantic Scholar metadata for
22	every paper that survived verification
23	- `workspace/refs.bib` — BibTeX file generated from the verified pool
24	- `workspace/drafts/intro_relwork.tex` — drafted Introduction and Related
25	Work sections, written into the template, with the rest of the template
26	preserved verbatim
27
28	## Two-phase pipeline (App. D.3)
29
30	```
31	PHASE 1 — Parallel Candidate Discovery
32	For each search direction in introduction_strategy.search_directions:
33	For each limitation_search_query in each related_work cluster:
34	- Use the host's web search tool to discover up to ~10 candidate papers.
35	- Run up to 10 discovery queries in parallel (host-permitting).
36	- Collect (title, snippet, url) tuples — no verification yet.
37	→ PRE-DEDUP before Phase 2 (see Step 1.5 below)
38
39	PHASE 2 — Sequential Citation Verification (1 QPS, with cache)
40	For each candidate (after pre-dedup), sequentially:
41	0. Check s2_cache.json first (scripts/s2_cache.py --check).
42	If HIT: use cached response, skip live S2 call. No throttle needed.
43	If MISS: proceed with live request below.
44	1. Query Semantic Scholar by title:
45	GET https://api.semanticscholar.org/graph/v1/paper/search?query=<title>
46	&fields=title,abstract,year,authors,venue,externalIds&limit=5
47	(Public endpoint, no key. Throttle to 1 QPS for live requests only.)
48	2. Store the S2 response in cache: s2_cache.py --store.
49	3. Pick the top hit. Check Levenshtein title ratio against the original
50	candidate title. If ratio < 70: discard.
51	4. Bonus: if year and venue exactly align with hints, add a +5 point
52	match-quality bonus.
53	5. Require: abstract is non-empty.
54	6. Require: paper.year (or month if known) strictly predates cutoff_date.
55	Months default to day-1: e.g., "October 2024" → 2024-10-01.
56	7. If all checks pass, add to verified pool.
57	After all candidates are verified, dedup by Semantic Scholar paperId.
58	```
59
60	The host agent does the LLM/web work; the deterministic helpers in `scripts/`
61	do the math.
62
63	## Step-by-step
64
65	### 0. Derive `cutoff_date`
66
67	Parse `conference_guidelines.md` for the submission deadline. The paper aligns
68	research cutoff with venue submission deadline (App. D.1):
69
70	\| Venue \| Cutoff \|
71	\|---\|---\|
72	\| CVPR 2025 \| Nov 2024 \|
73	\| ICLR 2025 \| Oct 2024 \|
74	\| Other \| One month before the stated submission deadline \|
75
76	Encode as `YYYY-MM-DD`. Months default to day-1 (e.g., `2024-10-01`).
77
78	### 1. Phase 1: Parallel Candidate Discovery
79
80	From `outline.json`:
81
82	- All `introduction_strategy.search_directions` (3-5 queries)
83	- For each cluster in `related_work_strategy.subsections`:
84	- The cluster's `sota_investigation_mission` becomes a search query
85	- All `limitation_search_queries` (1-3 each)
86
87	For each query, use your host's web search tool (e.g., `WebSearch` in
88	Claude Code, `@web` in Cursor, the search tool in Antigravity). Collect the
89	top ~10 candidates per query: title, abstract snippet, source URL.
90
91	If your host supports parallel sub-tasks, fire up to 10 concurrent search
92	queries. If not, run sequentially — slower but functionally equivalent.
93
94	#### Optional: Exa as a Phase 1 backend
95
96	If your host has no native web search, OR you want a research-paper-focused
97	backend with better signal-to-noise, you can use [Exa](https://exa.ai) via
98	the bundled `scripts/exa_search.py` helper. It is opt-in and reads
99	`EXA_API_KEY` from the environment — the repo never commits a key.
100
101	```bash
102	export EXA_API_KEY="your-key-here" # get one at https://dashboard.exa.ai/
103	python skills/literature-review-agent/s

People who liked this also viewed

6 picks

Type