Matrix Cognition

Paper Readings

What a paper claims, what it measured, and what holds up, cited by arXiv id.

Paper Readings take one recent arXiv paper at a time and ask three things: what it claims, what it actually measured, and what holds up. Every reading cites the paper by its arXiv id, names the authors and the date it was posted, and says where a table or a footnote qualifies the headline.

There are four so far. EarlyEval (arXiv 2609.02783) trains classifiers to stop agent benchmark runs once the outcome is predictable, and the reading checks what its Table III actually shows. Covert indirect prompt injection (arXiv 2608.30362) splits attack success on AgentDojo into outcomes the user notices and outcomes the user never sees, and shows the covert share can be engineered. The other two go further than reading. Where should a document live (arXiv 2609.17346v1) compares a document in the context window, in an injected KV cache and in the weights across five benchmarks, with its tables re-derived. TORQUE (arXiv 2609.36032v1) is re-implemented and re-run on Qwen2.5-0.5B: keeping outliers cut activation error at every budget, but the KV-cache gain faded once its side bits were paid for.

The readings do not depend on each other, so start with the one closest to your work. The TORQUE reading pairs with the quantisation piece in Systems, and the EarlyEval reading with Evaluation. The next reading in our queue is a recent paper on synthetic data or distillation.

  1. Paper Readings15 min read

    TORQUE re-implemented: where keeping outliers around a rotation pays

    A reading of arXiv 2609.36032v1, re-run on Qwen2.5-0.5B: keeping outliers cut activation error at every budget; the KV-cache gain faded once its bits were paid.

    arXiv 2609.36032CodeData

  2. Paper Readings8 min read

    Where should a document live: context, KV cache, or weights?

    A reading of arXiv 2609.17346v1, which compares in-context documents, KV-cache injection and fine-tuning on five benchmarks. Tables re-derived, code included.

    arXiv 2609.17346CodeData

  3. Paper Readings5 min read

    EarlyEval explained: stopping agent benchmark runs early

    A reading of arXiv:2609.02783, which trains classifiers to halt agent runs once the outcome is predictable, and what its Table III actually shows.

    arXiv 2609.02783

  4. Paper Readings6 min read

    Covert Indirect Prompt Injection (ICoA) explained

    A reading of arXiv:2608.30362, which splits attack success into covert and overt outcomes on AgentDojo and shows the covert share can be engineered.

    arXiv 2608.30362