Merge branch 'main' of github.com:cocoabench/cocoabench.github.io

Author: zzn-nzz

blog.html

@@ -174,9 +174,9 @@ <h4>Logical inference type</h4>
             For each task, we mark the <strong>dominant logical inference type</strong>: deductive, inductive, or abductive. Existing benchmarks typically concentrate on one direction (e.g., abductive reasoning in <a href="https://openai.com/index/browsecomp/">BrowseComp</a>, where the agent must infer a hidden target from partial clues). In contrast, CocoaBench deliberately spans all three, resulting in a more diverse set of reasoning questions.
         </p>
         <ul>
-            <li><strong>Deductive reasoning:</strong> from general rules to specific conclusions. For example, to solve the <a href="index.html#example-wb-logs">"W&B logs"</a> example task, the agent is given a set of specific rules and raw logs, and expected to reach a specific conclusion.</li>
-            <li><strong>Inductive reasoning:</strong> from multiple examples to a general pattern. For example, to solve the <a href="index.html#example-linear-regimes">"Linear regimes"</a> example task, the agent is given a set of raw data and expected to fit a function to the data.</li>
-            <li><strong>Abductive reasoning:</strong> from observed outcomes to the most plausible explanation or hypothesis. For example, to solve the <a href="index.html#example-8-puzzle-game">"8-puzzle game"</a> example task, the agent is needs to infer the target state based on interaction with the interface.</li>
+            <li><strong>Deductive reasoning:</strong> from general rules to specific conclusions. For example, in the <a href="index.html#example-wb-logs">"W&B logs"</a> example task, the agent is given specific rules and raw logs, and expected to reach a specific conclusion.</li>
+            <li><strong>Inductive reasoning:</strong> from multiple examples to a general pattern. For example, in the <a href="index.html#example-linear-regimes">"Linear regimes"</a> example task, the agent is given raw data and expected to fit a function to the data.</li>
+            <li><strong>Abductive reasoning:</strong> from observed outcomes to the most plausible explanation or hypothesis. For example, to solve the <a href="index.html#example-8-puzzle-game">"8-puzzle game"</a> example task, the agent needs to infer the target state based on interaction with the interface.</li>
         </ul>
         <p>
             Typically, tasks are mixed with all different reasoning directions. We do not treat these categories as rigid boxes. Instead, they provide a <strong>lens for analysis</strong>: for each task, we mark which directionality is most important, allowing us to ask, for example, whether agents struggle more with abductive tasks (explaining failures) than with purely deductive ones (checking constraints).