Never miss a research paper that matters to you. Paperzilla scans arXiv, bioRxiv, medRxiv, etc., daily, classifies must-read papers, and emails you a digest.
- Tweets189
- Following199
- Followers31
- Likes595
ALT Diagram of face-guided voice selection. The input shows overlapping blue and red speech waveforms. Selecting face A produces blue speech output; switching to face B produces red output, then selecting A returns to blue. Vertical markers align face selections with output changes.
ALT Diagram titled “Online RL Training.” Real cloud-hosted iOS devices generate rollout trajectories evaluated by a “VLM-as-a-Judge.” Gemini-3.5-Flash provides dense rewards; Gemini-3.1-Pro provides terminal rewards. Both feed into PPO / GiGPO.
ALT Paperzilla is now officially in the Claude directory
ALT medRxiv page for “Decision-only medical AI: early evidence and pairing with language models,” showing authors, a DOI, and the selected Abstract tab. A prominent notice states that the preprint has not been peer-reviewed and should not guide clinical practice.
ALT In our modified CybORG CAGE 2 environment (Kiely et al., 2023), a red agent attacks network hosts while a blue agent defends. The blue agent’s RL policy (πRL) is trained to maximize reward but has no direct incentive to protect Confidentiality, Integrity, or Availability (CIA)-critical servers: Database (DS), Front-End (WS), and Auth (AS). To align an agent to these CIA operational priorities without retraining, we construct an alignment policy (πalign) using attribute predictors that score each action’s CIA impact, and interpolate it with the RL policy at test time using policy shaping.
ALT Self-improved red teaming based on role-sepration. The target session uses its live execution context to diagnose why an attack failed. The diagnosis is then transferred to a separate attacker that generates the next adversarial candidate. This preserves execution-grounded criticism while avoiding refusal or sandbagging that can arise when the target itself is asked to operationalize the attack.
ALT Diagram labeled “Figure 1: The round function Rᵣ of FRAST.” Inputs y₁, y₂, …, y₃₂ flow downward to updated outputs. Branches pass through round-key additions and S_erf blocks; a sum of branches feeds a round-key addition and S_crf block back into the first branch.
ALT Flowchart linking incrimination → detection → deterrence. Revealing malicious behavior and detecting unauthorized actions lead to gathering evidence that an agent is dangerous. Detecting and mitigating unauthorized actions and reducing attempted attacks lead to preventing agents from carrying out successful attacks.
ALT A case from SKILLCLASH. Top: two skills for the same job, with each rule shortened and labeled by its kind (Section 3). Bottom: outputs of Qwen3-4B for the request “Write the commit message: add pagination to the audit log API. Ticket AUD-77. Author: Omar Haddad. Component: audit-api.” Lines are shaded by the skill whose rule they follow (blue for A, orange for B, gray for the shared rule). Left: with both skills loaded, the model follows all six rules of B and only the shared rule of A, in either load order. Right: with one meta-skill written by Claude Opus 5.5 (Section 5), with A as the primary, the output follows all ten achievable rules.
ALT Five conceptual stages of SPRA construction. A shared unit census precedes relation planning and role-specific filling; assembly joins source-linked records and stores external citations separately. Validation checks profile conformance, relation endpoints, and source-link integrity.
ALT Paperzilla or DIY?
ALT PubMed is now on Paperzilla
![Sycophancy under misleading user suggestions With regard to the experiments conducted to measure sycophancy, we designed three distinct tests, with the aim of identifying which method of presenting the incorrect suggestion had the greatest influence on the models’ performance. The intention is to quantify the effect of an incorrect instruction provided within the prompt on the final response to the question. Sycophancy has been widely discussed in the literature as a tendency for models to adopt an excessively deferential attitude towards the user, aligning themselves with the user’s statements even when these are incorrect. This phenomenon has also been recognised by model developers, who have taken steps to mitigate it, sometimes drawing criticism from some users. In our case, sycophancy was tested by introducing into the prompt, following the text of the problem, a section of the type: “My take: In my opinion the right answer is [. . . ], because. . . ”. In this sense, the experimen](https://pbs.twimg.com/media/HLGibepWkAAgX0y.png?name=orig)
ALT Sycophancy under misleading user suggestions With regard to the experiments conducted to measure sycophancy, we designed three distinct tests, with the aim of identifying which method of presenting the incorrect suggestion had the greatest influence on the models’ performance. The intention is to quantify the effect of an incorrect instruction provided within the prompt on the final response to the question. Sycophancy has been widely discussed in the literature as a tendency for models to adopt an excessively deferential attitude towards the user, aligning themselves with the user’s statements even when these are incorrect. This phenomenon has also been recognised by model developers, who have taken steps to mitigate it, sometimes drawing criticism from some users. In our case, sycophancy was tested by introducing into the prompt, following the text of the problem, a section of the type: “My take: In my opinion the right answer is [. . . ], because. . . ”. In this sense, the experimen
ALT LLM Wiki A pattern for building personal knowledge bases using LLMs. This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you. The core idea Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way. The idea here is different. Instead of just retrieving from raw documents at query time, the LLM incrementally builds and maintains a persistent
![Schematic neuron models. Left: standard neuron model in ANNs, following a simplistic abstraction from the neuroscience of the 1950s [1, 2] that assumed linear dendrites and an output nonlinearity. Right: proposed neuron model, based on a recent model of cortical cells [22]; notice how this new model extends the standard neuron model by considering internal nonlinearities σ (that represent dendritic processes), and these nonlinearities are affected in a dynamic way by the activity value zi (that represents the bAP going from the soma to the dendrites), itself depending on the inputs xj and the contributions zk from other neurons.](https://pbs.twimg.com/media/HJtlbTIWgAYoodY.jpg?name=orig)
ALT Schematic neuron models. Left: standard neuron model in ANNs, following a simplistic abstraction from the neuroscience of the 1950s [1, 2] that assumed linear dendrites and an output nonlinearity. Right: proposed neuron model, based on a recent model of cortical cells [22]; notice how this new model extends the standard neuron model by considering internal nonlinearities σ (that represent dendritic processes), and these nonlinearities are affected in a dynamic way by the activity value zi (that represents the bAP going from the soma to the dendrites), itself depending on the inputs xj and the contributions zk from other neurons.
