Claude Haiku
No mentions found
This entity hasn't been tracked yet, or Iris is still building its knowledge base.
Related Articles from SNS
Scaffold Effects on GAIA: A Controlled Comparison
Announce Type: new Abstract: Published agent capability scores conflate what a model can do with what its scaffold lets it do, and the magnitude of this elicitation gap is not well characterized under controlled conditions. This study executes a pre-registered controlled comparison of three scaffolds (ReAct, a Planner-Actor-Rater multi-agent design, and planner-then-executor) across five models from three providers (Claude Opus 4.7, Sonnet 4.6, Haiku 4.5; Gemini 3.1 Pro Preview; GPT-5.5) on...
Long Live Fine-Tuning: Task-Specific Transformers Outperform Zero-Shot LLMs for Misinformation Response Classification on Reddit
Announce Type: new Abstract: As large language models (LLMs) become default tools for online information verification, an implicit assumption follows them: that scale and general capability are sufficient for nuanced classification of misinformation discourse. We test this assumption directly on 900 Reddit comments spanning three PolitiFact-verified misinformation claims (environment, health, immigration), labelled as belief (propagates the claim), fact-check (corrects it), or other. We...
Depth-Dependent Indirect Prompt Injection in Tool-Calling ReAct Agents: Injection Depth, Payload Framing, and Turn-Budget Sensitivity
Announce Type: new Abstract: ReAct agents that interleave chain-of-thought reasoning with tool calls are increasingly deployed for real tasks such as scheduling, file retrieval, and data access. Their tool observation loop creates a direct attack surface: an adversary who controls any tool's return value can embed instructions that redirect the agent away from the user's goal, a threat known as indirect prompt injection. Existing benchmarks evaluate attack success rate (ASR) at a fixed...
When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agents
arXiv:2602.08235v2 Announce Type: replace Abstract: Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this...
Microsoft tells engineers to stop using Anthropic's Claude
Microsoft is canceling most internal Claude Code licenses by June 30, pushing engineers to its own GitHub Copilot CLI. Anthropic's tool got too popular, undercutting Microsoft's homegrown product. The official reason is toolchain unification, but The Verge reports the fiscal-year-end timing points to cost-cutting.
Ask HN: What are tools you have made for yourself since the advent of AI?
I've made a number of ceramic molds for slumping fused glass into bowls. As well as wooden templates for ceramic mugs. I've devised a few carrying tools to move glass frit paintings from my studio down to my barn where the kilns sit without spilling the glass.
Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
Hey HN, we’re Ismaeel, Eren, Yafet and Nikodem. We built Expanse (https://expanse.sh/) to increase the effective capacity of your HPC/GPU clusters running schedulers/orchestrators like Kubernetes and SLURM. We read the source code, job submission script, and the hardware a workload is about to run on to predict what the job actually needs before the cluster sees it.
TriEval: A Resource-Efficient Pipeline for LLM Bias, Toxicity, and Truthfulness Assessment
arXiv:2606.03036v1 Announce Type: new Abstract: LLMs have evolved from basic chatbots to the backbone of the AI ecosystem, now widely used in healthcare, schools, and government services. The domain-wide adoption of LLMs necessitates continuous evaluation to ensure their safety and fairness. Common issues encountered after deploying LLMs include inconsistent outputs and hallucinations of incorrect information.
From `May' to `Is': Certainty Distortion in Language Model Rewriting
arXiv:2606.07951v1 Announce Type: new Abstract: Humans increasingly turn to Language Models (LMs) in ways that shape beliefs and drive decisions, including discussing, rewriting, and summarizing information from scientific articles, news, and medical reports. However, in these domains, where how confidently a claim is expressed matters, little is known about whether LMs faithfully preserve it. In this work, we investigate certainty distortion in LMs, defined as meaningful changes in...
OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification
arXiv:2606.01476v1 Announce Type: new Abstract: On-Policy Distillation (OPD) trains a student model on its own generative trajectories under dense token-level feedback from a stronger teacher, mitigating both the off-policy distribution shift of Supervised Fine-Tuning (SFT) and the sparse credit assignment of Reinforcement Learning (RL). However, standard OPD faces two coupled limitations. First, it requires direct access to the teacher's token-level logits, excluding a broad class of...