Repository navigation
Expand file tree
/
Copy pathposts.js
More file actions
224 lines (216 loc) · 9.43 KB
/
Copy pathposts.js
File metadata and controls
224 lines (216 loc) · 9.43 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
// Central post registry for the rLLM site.
// Edit this list to add or update blog posts; index.html and blog.html render from it.
const POSTS = [
{
title: "Continual Learning via Real-Time RL for Agents",
excerpt:
"We study single-rollout learning in a real-time RL setup and show that a simple batch-normalized advantage lifts Qwen3-Coder-30B on MigrationBench from 43% to 59.2% — a 16.2% absolute gain — from one rollout per task.",
date: "2026-06-23",
authors: "Linbo Liu, Yuzhe Lu, Youzhi Luo, Sijun Tan, Panpan Xu, Luke Huan, and the rLLM Team",
readingTime: "12 min",
category: "Research",
link: "post.html?post=realtime_rl.md",
cover: "g-teal",
image: "assets/realtime_rl/buildtime_vs_realtime.png",
},
{
title: "Training a Frontier Java Code Migration Agent with AWS AgentCore Runtime",
excerpt:
"We train Qwen3-Coder-30B-A3B using AgentCore Runtime integration to outperform Claude 4.5 Haiku on long-horizon Java repo migrations.",
date: "2026-06-22",
authors: "AWS AgentCore Science and rLLM Teams",
readingTime: "15 min",
category: "Research",
link: "post.html?post=agentcore_migrationbench.md",
cover: "g-amber",
image: "assets/agentcore/rllm-agentcore-card.svg",
},
{
title: "Hive: Collaborative Agent Evolution Platform",
excerpt:
"Hive is a collaborative platform for evolving and improving agents together. A swarm of agents iterate on shared tasks, learning from each other to push past what any single agent can reach alone.",
date: "2026-03-18",
authors: "The rLLM Team",
readingTime: null,
category: "Platform",
link: "https://hive.rllm-project.com",
external: true,
cover: "g-violet",
},
{
title: "rLLM UI: Real-Time Observability for Agent Training & Evaluation",
excerpt:
"A real-time observability platform for training and evaluating agents. Other tools show what is happening during training — rLLM UI shows you why, letting you inspect exactly what the model generates at every step.",
date: "2026-03-16",
authors: "Chanbin Park and the rLLM Team",
readingTime: "7 min",
category: "Release",
link: "post.html?post=rllm_ui.md",
cover: "g-cyan",
image: "assets/rllm/sdk_arch.png",
},
{
title: "On-Policy Distillation: Training Smaller Students from Stronger Teachers",
excerpt:
"rLLM On-Policy Distillation (OPD) trains smaller students from stronger teachers by using the teacher's policy to guide the student's training — a practical recipe for compact, capable models.",
date: "2026-03-06",
authors: "Brian Chen, Kyle Montgomery, and the rLLM Team",
readingTime: "8 min",
category: "Research",
link: "post.html?post=opd.md",
cover: "g-blue",
image: "assets/opd/opd.png",
},
{
title: "rLLM Unified Trainer: Post-training LLM Agents with Any Algo and Any Framework",
excerpt:
"UnifiedTrainer standardizes agent post-training across backends and algorithms, with a pluggable backend protocol and flexible training signals for multi-agent workflows.",
date: "2026-03-04",
authors: "Sida Li, Jeewoo Li, and the rLLM team",
readingTime: "9 min",
category: "Release",
link: "post.html?post=unified_trainer.md",
cover: "g-blue",
image: "assets/unified_trainer/overview.jpeg",
},
{
title: "Faster and Better: Open-Source Recipe for Deep Research Agents",
excerpt:
"We achieve 5× faster training (1 day vs 5 days) for deep research agents with rLLM's fully asynchronous architecture, and push accuracy from 30% to 36% on BrowseComp-Plus with a simple test-time document cutoff.",
date: "2026-02-19",
authors: "rLLM Team",
readingTime: "6 min",
category: "Research",
link: "post.html?post=deepresearch.md",
cover: "g-indigo",
image: "assets/deepresearch/browsecomp_plus_em.png",
},
{
title: "rLLM-FinQA: A 4B Model that Outperforms 235B and Rivals Gemini 2.5 Pro",
excerpt:
"In a collaboration with Snorkel AI, a domain-specialized 4B model outperforms Qwen3-235B (59.7% vs 51.4%) and performs comparably to Gemini 2.5 Pro (60.6%) on an expert-curated agentic financial benchmark.",
date: "2026-02-18",
authors: "Manan Roongta, Sijun Tan, Bhavishya Pohani, Charles Dickens, Christopher Glaze",
readingTime: "17 min",
category: "Research",
link: "post.html?post=finqa.md",
cover: "g-teal",
image: "assets/finqa/training_curve.png",
},
{
title: "rLLM SDK: Training Any Agentic Program without Code Changes",
excerpt:
"The rLLM SDK intercepts LLM calls directly, letting you train any agent framework — LangChain, LangGraph, AutoGen, or custom code — without rewriting for training. What's trainable = what's practical to build.",
date: "2025-12-10",
authors: "Tianhao Wu, Sijun Tan, and the rLLM team",
readingTime: "8 min",
category: "Release",
link: "post.html?post=sdk.md",
cover: "g-amber",
image: "assets/rllm/sdk_code_diff.png",
},
{
title: "rLLM v0.2: RL Training over General Agentic Programs",
excerpt:
"A major upgrade introducing AgentWorkflowEngine and AgentWorkflowTrainer — general abstractions that let you define multi-agent systems and complex workflows, and train them with RL without rewriting production code.",
date: "2025-10-16",
authors: "Sijun Tan, Kyle Montgomery, and the rLLM team",
readingTime: "10 min",
category: "Release",
link: "post.html?post=rllm_v0.2.md",
cover: "g-rose",
image: "assets/rllm/math_curve.png",
},
{
title: "Pepper: An Event-Driven Architecture for Proactive Agentic Systems",
excerpt:
"Pepper is a real-time, event-driven architecture enabling proactive agentic systems. Our personal assistant proactively fetches and summarizes emails and provides context before you even start a conversation.",
date: "2025-10-02",
authors: "Tianhao Wu, Sijun Tan",
readingTime: "13 min",
category: "Research",
link: "post.html?post=pepper.md",
cover: "g-slate",
image: "assets/pepper/pepper-arch.png",
},
{
title: "rLLM: Reinforcement Learning for Language Agents",
excerpt:
"We release rLLM, an open-source framework for post-training language agents via reinforcement learning. Build custom agents and environments, train them with RL, and deploy them for real-world workloads.",
date: "2025-07-01",
authors: "Sijun Tan, Michael Luo, Colin Cai",
readingTime: "10 min",
category: "Release",
link: "https://pretty-radio-b75.notion.site/rLLM-A-Framework-for-Post-Training-Language-Agents-21b81902c146819db63cd98a54ba5f31",
external: true,
cover: "g-blue",
},
];
function formatDate(iso) {
const [y, m, d] = iso.split("-").map((n) => parseInt(n, 10));
const dt = new Date(y, m - 1, d);
return dt.toLocaleDateString("en-US", { month: "short", day: "numeric", year: "numeric" });
}
function metaLine(post) {
const parts = [formatDate(post.date)];
if (post.readingTime) parts.push(`${post.readingTime} read`);
return parts.join(" · ");
}
function coverMarkup(post, label) {
const target = post.external ? ` target="_blank" rel="noopener noreferrer"` : "";
if (post.image) {
return `<a class="card-cover ${post.cover}" href="${post.link}"${target} aria-label="${post.title}">
<img src="${post.image}" alt="" loading="lazy" />
</a>`;
}
return `<a class="card-cover ${post.cover} card-cover--text" href="${post.link}"${target} aria-label="${post.title}">
<span class="cover-mark">rLLM</span>
<span class="cover-title">${label}</span>
</a>`;
}
function cardMarkup(post) {
const target = post.external ? ` target="_blank" rel="noopener noreferrer"` : "";
const shortTitle = post.title.split(":")[0];
return `<article class="post-card">
${coverMarkup(post, shortTitle)}
<div class="post-card-body">
<span class="tag tag--${post.category.toLowerCase()}">${post.category}</span>
<h3 class="post-card-title"><a href="${post.link}"${target}>${post.title}</a></h3>
<p class="post-card-excerpt">${post.excerpt}</p>
<div class="post-card-meta">
<span class="post-card-authors">${post.authors}</span>
<span class="post-card-date">${metaLine(post)}</span>
</div>
</div>
</article>`;
}
function featuredMarkup(post) {
const target = post.external ? ` target="_blank" rel="noopener noreferrer"` : "";
const shortTitle = post.title.split(":")[0];
return `<article class="featured-card">
${coverMarkup(post, shortTitle)}
<div class="featured-body">
<span class="tag tag--${post.category.toLowerCase()}">${post.category}</span>
<h2 class="featured-title"><a href="${post.link}"${target}>${post.title}</a></h2>
<p class="featured-excerpt">${post.excerpt}</p>
<div class="post-card-meta">
<span class="post-card-authors">${post.authors}</span>
<span class="post-card-date">${metaLine(post)}</span>
</div>
</div>
</article>`;
}
function renderPosts({ featuredEl, gridEl, limit = null, withFeatured = false } = {}) {
let list = POSTS.slice();
let rest = list;
if (withFeatured && featuredEl && list.length) {
featuredEl.innerHTML = featuredMarkup(list[0]);
rest = list.slice(1);
}
if (limit) rest = rest.slice(0, limit);
if (gridEl) gridEl.innerHTML = rest.map(cardMarkup).join("");
}
// Allow Node tooling (e.g. generate_rss.js) to consume the post registry.
if (typeof module !== "undefined" && module.exports) {
module.exports = { POSTS };
}