You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 3f2eeb6
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: README.md
+2Lines changed: 2 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -15,6 +15,8 @@ node website/check.mjs
15
15
python3 -m http.server 4174 --bind 127.0.0.1
16
16
```
17
17
18
+
The page opens with the full paper title. Each figure has a reading guide; benchmark setups describe the parallel fan-in and sequential chain. The booking guide distinguishes selected source positions from generated text messages.
19
+
18
20
The page shows the paper’s Figure 2 beside a short TL;DR, both benchmark plot
19
21
groups, an open booking replay, and an always-visible citation. The booking replay
20
22
comes before the TL;DR; the coding video follows the results. Copy icons briefly confirm success before fading back.
<h2id="hero-results-title" class="eyebrow">FANOUTQA</h2><pclass="visual-note">Three senders read different Wikipedia pages; one receiver combines their messages to answer a question. Each row compares same-size models: higher accuracy and lower time are better. pp = percentage points.</p>
<divclass="booking-heading"><div><pclass="eyebrow">Interactive demo</p><h2id="booking-question">Where should Maya collect her pass on Friday, and what ID should she bring?</h2></div><aclass="text-link" href="demo/index.html" target="_blank" rel="noopener noreferrer">Open full window <spanaria-hidden="true">↗</span></a></div>
65
+
<pclass="visual-note">Press <strong>Run</strong> to replay two Qwen3-8B chains. Agents 1-3 each read a different document and pass a message to the next agent; Agent 4 answers. <strong>Top:</strong> CacheBack highlights source positions whose state is selected for the next agent, guided by the question. Highlights change at each handoff as inherited and new positions are selected. <strong>Bottom:</strong> text messages appear as they are generated. The last column shows each chain’s answer. Timers show recorded elapsed time; streaming is reconstructed.</p>
66
66
<divclass="recorded-demo">
67
67
<divclass="demo-window"><iframeid="flagship-demo" src="demo/index.html?v=fit-7" title="Recorded CacheBack and text booking relay comparison, with playback controls" loading="lazy" sandbox="allow-scripts allow-same-origin" referrerpolicy="no-referrer"></iframe></div>
<figureclass="overview-figure"><ahref="https://arxiv.org/pdf/2609.32046" aria-label="Read Figure 2 in the paper"><imgsrc="website/assets/figures/communication-overview.webp" width="1600" height="393" loading="lazy" alt="Figure 2: sending every agent's full cache exceeds the receiver's context limit. CacheBack sends selected positions and latent steps, leaving room for the receiver's prompt."></a><figcaption>Figure 2. Select what the receiver needs, instead of accumulating every sender’s context.</figcaption></figure>
72
-
<divclass="tldr"><h2id="abstract-title">TL;DR</h2><p>Agents can share <strong>internal model state</strong> instead of writing reports. Sending all of it, however, overwhelms the receiver.</p><p><strong>The receiver asks for what it needs.</strong> CacheBack uses attention to that request to select which state gets sent. Simple, robust, and <strong>training-free.</strong></p></div>
71
+
<figureclass="overview-figure"><ahref="https://arxiv.org/pdf/2609.32046" aria-label="Read Figure 2 in the paper"><imgsrc="website/assets/figures/communication-overview.webp" width="1600" height="393" loading="lazy" alt="Figure 2: sending every agent's full cache exceeds the receiver's context limit. CacheBack sends selected positions and latent steps, leaving room for the receiver's prompt."></a><figcaption>Figure 2. Full KV caches (stored model state) accumulate source context at the receiver and can exceed its context limit. Receiver-conditioning selects source and latent positions, leaving room for the receiver’s prompt. Token IDs represent source positions; continuous vectors represent latent steps.</figcaption></figure>
72
+
<divclass="tldr"><h2id="abstract-title">TL;DR</h2><p>Agents can share <strong>internal model state</strong> instead of generating text messages. Sending all of it, however, overwhelms the receiver.</p><p><strong>The receiver asks for what it needs.</strong> CacheBack uses attention to that request to select which state gets sent. Simple, robust, and <strong>training-free.</strong></p><p>The title’s 94% refers to removing 93.75% of sender positions at 16× compression.</p></div>
<divclass="curves-heading"><h3>FanOutQA</h3><pid="compression-guide">Three senders read separate sources; one receiver answers. <strong>Higher is more accurate; left is faster.</strong> CacheBack points vary message size; text points vary sender size. <strong>rX keeps 1/X of the sender’s positions: r4 keeps 25%.</strong></p></div>
77
+
<divclass="curves-heading"><h3>FanOutQA</h3><pid="compression-guide">Questions require combining evidence across Wikipedia pages. We split at least 120K tokens among three parallel senders; a receiver answers from their messages. Strict accuracy requires every reference-answer group. <strong>Higher is more accurate; left is faster.</strong> Each panel fixes the receiver model. Text labels give sender size; CacheBack uses same-size senders. <strong>rX retains 1/X of sender positions; r4 keeps 25%.</strong></p></div>
78
+
<divclass="topology-note fan-in"><div><span>Pages A → Sender 1 ↘</span><span>Pages B → Sender 2 →</span><span>Pages C → Sender 3 ↗</span></div><strong>Receiver → Answer</strong></div>
78
79
<divclass="curve-legend" aria-label="Curve legend"><spanclass="rclc-emphasis">● ━ CacheBack</span><spanclass="text-emphasis">▲ ┄ Text · labels show sender size</span><span>× Question only</span></div>
79
80
<divclass="curve-grid">
80
-
<figure><imgsrc="website/assets/figures/fanoutqa-qwen.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="FanOutQA, Qwen 3 8B: CacheBack r4 reaches 55.3% strict accuracy at 192 seconds, compared with 40.7% at 612 seconds for same-size text; weaker settings are also plotted."></figure>
81
-
<figure><imgsrc="website/assets/figures/fanoutqa-nemotron.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="FanOutQA, Nemotron Nano 2 12B: CacheBack r8 reaches 50% accuracy near 102 seconds versus 38.7% near 131 seconds for same-size text. A smaller text sender is faster at lower accuracy."></figure>
81
+
<figure><imgsrc="website/assets/figures/fanoutqa-qwen.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="FanOutQA, Qwen 3 8B: CacheBack r4 reaches 55.3% strict accuracy at 192 seconds, compared with 40.7% at 612 seconds for same-size text; weaker settings are also plotted."><figcaption>Qwen: r4 improves accuracy and completion time over 8B text; 1.7B text is slightly faster but less accurate.</figcaption></figure>
82
+
<figure><imgsrc="website/assets/figures/fanoutqa-nemotron.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="FanOutQA, Nemotron Nano 2 12B: CacheBack r8 reaches 50% accuracy near 102 seconds versus 38.7% near 131 seconds for same-size text. A smaller text sender is faster at lower accuracy."><figcaption>Nemotron: r8 improves on 12B text; 4B text is faster but less accurate.</figcaption></figure>
82
83
</div>
83
-
<divclass="chain-results" id="chain-results"><divclass="curves-heading"><h3>LongBench v2 Easy</h3><p>Four agents read a long document in sequence, passing state to a final answering agent. Relative budgets (rX) grow with context; fixed budgets (32K, 64K) cap message positions. Stronger compression can reduce accuracy.</p></div>
84
+
<divclass="chain-results" id="chain-results"><divclass="curves-heading"><h3>LongBench v2 Easy</h3><p>Long-document questions test whether evidence survives repeated handoffs. We split each 100K-246K-token document into four equal parts. Each sender reads its part plus the previous message; a final receiver answers from the fourth message. Relative budgets (rX) grow with accumulated context; fixed budgets (32K, 64K) cap message positions. Axes and text-size labels follow FanOutQA.</p></div>
85
+
<pclass="topology-note">Part 1 → Agent 1 → Part 2 + Agent 2 → Part 3 + Agent 3 → Part 4 + Agent 4 <strong>→ Receiver → Answer</strong></p>
<figure><imgsrc="website/assets/figures/longbench-qwen.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="LongBench v2 Easy, Qwen 3 8B: both relative and fixed CacheBack budgets include settings faster and more accurate than same-size text. Stronger compression eventually lowers accuracy."></figure>
87
-
<figure><imgsrc="website/assets/figures/longbench-nemotron.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="LongBench v2 Easy, Nemotron Nano 2 12B: CacheBack r8 reaches 48% accuracy near 133 seconds, while fixed-budget settings show a different tradeoff; smaller text senders can be faster."></figure>
88
+
<figure><imgsrc="website/assets/figures/longbench-qwen.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="LongBench v2 Easy, Qwen 3 8B: both relative and fixed CacheBack budgets include settings faster and more accurate than same-size text. Stronger compression eventually lowers accuracy."><figcaption>Qwen: relative r4 gives the highest plotted accuracy; stronger compression loses accuracy.</figcaption></figure>
89
+
<figure><imgsrc="website/assets/figures/longbench-nemotron.svg?v=sender-sizes" width="560" height="460" loading="lazy" alt="LongBench v2 Easy, Nemotron Nano 2 12B: CacheBack r8 reaches 48% accuracy near 133 seconds, while fixed-budget settings show a different tradeoff; smaller text senders can be faster."><figcaption>Nemotron: relative r8 beats 12B text on both axes; 4B text remains faster but less accurate.</figcaption></figure>
88
90
</div>
89
91
</div>
90
-
<pclass="curve-caption">50 concurrent tasks on 8 H100s; accuracy averaged over three receiver draws. Completion includes queueing. <ahref="https://arxiv.org/pdf/2609.32046">Full setup and results in the paper.</a></p>
92
+
<pclass="curve-caption">Lines join the best accuracy-time tradeoffs within each channel; faded points are dominated settings. × means the receiver gets only the question. 50 concurrent tasks on 8 H100s; accuracy averaged over three receiver draws. Completion includes queueing. <ahref="https://arxiv.org/pdf/2609.32046">Full setup and results in the paper.</a></p>
<divclass="demo-description"><div><pclass="demo-type">36 seconds · 4K · Recorded coding task</p><h3id="coding-title">Fixing a Django bug with seven coding agents</h3><p>Qwen3-8B produces the patch <strongclass="rclc-emphasis">4.41× faster</strong>with CacheBack than with written reports on this case. Both runs make the same one-line code change and pass all 88 tests.</p></div></div>
99
+
<divclass="demo-description"><div><pclass="demo-type">36 seconds · 4K · Recorded coding task</p><h3id="coding-title">Fixing a Django bug with seven coding agents</h3><p>Seven Qwen3-8B workers inspect the code and send information to a coordinator, which writes the Django fix. The video compares selected latent state with generated text messages. CacheBack produces the patch <strongclass="rclc-emphasis">4.41× faster</strong> than with text communication on this case. Both runs make the same one-line code change and pass all 88 tests.</p></div></div>
<p><ahref="demo/coding/rclc-cacheback-4k.mp4">Watch the demo as an MP4</a>.</p>
101
103
</video><buttonclass="play-demo" id="play-demo" type="button" hidden><spanaria-hidden="true">▶</span> Watch the full demo</button></div>
102
-
<divclass="demo-caption" id="coding-caption"><p><strong>One recorded case:</strong> CacheBack 25.66 s; text 113.21 s. Startup and test grading excluded. Silent video; playback speed varies. <ahref="demo/coding/rclc-cacheback-4k.mp4">Open in full window</a></p>
104
+
<divclass="demo-caption" id="coding-caption"><p><strong>One recorded case:</strong> CacheBack 25.66 s; text 113.21 s. Separate recorded runs are aligned at their start. Startup and test grading excluded. Silent video; playback speed varies. <ahref="demo/coding/rclc-cacheback-4k.mp4">Open in full window</a></p>
0 commit comments