Wednesday, August 5, 2026

1. Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition
   T2arXiv AI/CL/LG（论文）

   This study explores linear representations of temporal horizon in the large language model Qwen3-32B, identifying short-term vs long-term directions in the residual stream using contrastive linear probes. Contrastive activation addition (CAA) steering is evaluated on a held-out binary temporal-choice task, an out-of-distribution monetary intertemporal-choice task, and the TravelPlanner benchmark. Results show large bidirectional shifts in the model's indifference threshold between smaller-sooner and larger-later rewards, and moderate steering improves planning-related capability metrics. The findings indicate that model intertemporal preferences are measurable and steerable, with implications for AI systems giving advice on delayed costs and benefits, as well as safety concerns around long-horizon planning.

2. llm-anthropic 0.26 Released: Adds Claude 5 Models and Server-Side Tool Support
   T2Simon Willison 博客（Atom）

   llm-anthropic 0.26 is released, leveraging LLM 0.32 to add Claude Fable 5, Sonnet 5, and Opus 5 models, along with server-side tools such as WebSearch, WebFetch, CodeExecution, and AnthropicMCP, callable via LLM's -T interface or Python tools=. Legacy -o web_search* options are removed in favor of -T WebSearch. Reasoning, tool calls, and results now stream as typed events, with CLI reasoning displayed on stderr unless hidden via -R. Claude 5 models think by default, with thinking_effort control, and Fable 5 always thinks. The update simplifies extended thinking configuration and improves the model integration experience.

3. Can Large Language Models Recover Semantic Optimization Opportunities That Compilers Miss?
   T2arXiv AI/CL/LG（论文）

   This paper introduces SeGaBench, a benchmark to test whether large language models (LLMs) can recover semantic information missing from C/C++ code and generate validated optimization artifacts. Across 120 executable cases, the strongest model produced correct artifacts in 94.8% of responses, achieved at least 1.05x speedup in 83.3%, and attained performance success on 93.3% of cases. However, correct artifacts often only closed part of the oracle gap. The study shows that LLMs can complement compiler analysis as speculative semantic proposers, provided their outputs are validated and evaluated.

4. Open-weight models approach frontier capability, safety gap remains: SaferAI report highlights GLM-5.2
   T2TechCrunch：AI（RSS）

   A new SaferAI report finds that Z.ai's open-weight model GLM-5.2 approaches frontier AI capabilities on multiple benchmarks, yet lacks critical safety mitigations. This renews concerns that powerful open models may outpace governance and safeguards, resulting in a widening safety gap that is difficult to control. The report calls on developers and policymakers to pay closer attention to deployment risks of open-weight models and to create proper assessment and protective mechanisms before release. It also emphasizes the need for transparency and independent audits as open models increasingly become part of mainstream AI applications.

5. MiniMax-H3 Ported to MLX, Runs on Apple Silicon to Generate Videos with Audio
   T2Simon Willison 博客（Atom）

   Simon Willison blogged about PipeNetwork/minimax-h3-mlx, a port of MiniMax-H3, released two days ago as a general-purpose omni-modal generative system, to the MLX framework for Apple Silicon. He tested it on an M5 Max MacBook Pro, downloading ~115GB of model files and generating a 15-second video with audio in just under 45 minutes. The video is impressive, but the audio sounds like weird speech-like garbage because no prompt guidance was provided for audio. He notes that the prompting guide offers tips for better results.

1. LLM 0.32 Released: Reasoning Traces, OpenAI Responses, and Server-Side Tools
   T2Simon Willison 博客（Atom）

   Simon Willison released LLM 0.32, the most significant upgrade since the project's initial launch. The new version adds visible reasoning traces for reasoning models, support for the OpenAI Responses API, server-side tools from providers, and redesigned content-addressable SQLite logs. The default model is now GPT-5.6 Luna, an inexpensive but capable option. Python API users get a new messages parameter and a stream_events interface that can handle reasoning text, tool calls, and attachments. New commands like llm openai endpoint enable one-liner prompts against any OpenAI-compatible service, while plugins for Anthropic, Gemini, and OpenRouter also received substantial updates. With these changes, LLM is beginning to look like an agent framework, capable of mixing tools from different sources with different models.

2. Claude Code v2.1.222 Released: Fixes Worktree Isolation, Proxy Connectivity, and Stability Issues
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.222 has been released, fixing a vulnerability that allowed worktree-isolated sessions and their subagents to run destructive git commands against the main checkout; isolation now applies to file edits and Bash in every session type. It also fixes PreToolUse auto-allow hooks bypassing tool restrictions in background agent tasks, and resolves startup connectivity checks hanging behind HTTPS proxies by using the same proxy-aware transport as API requests. Other improvements include truncating long SendMessage summaries, correcting usage attribution for MCP servers, adjusting Remote Control auto-start so repo-local settings cannot enable it, and removing the ultraplan feature. This release enhances security, stability, and user experience.

3. Memcode Launches Terminal Coding Agent with Persistent Codebase Memory
   T2Hacker News 热门（Algolia · 中文化）

   Memcode is a new agentic coding CLI that builds a persistent model of your codebase to provide better context and memory across sessions. With a one-line installation command, developers can quickly deploy an AI agent in their terminal to understand, navigate, and modify large project code more efficiently. The tool targets programmers who want a higher level of automation in their daily workflow by reducing the need to repeatedly reload context. It represents an emerging trend of terminal-native AI agents that remember the structure and logic of complex codebases, aiming to boost continuity and accuracy in coding tasks. Early adopters and AI coding tool enthusiasts may find it valuable to test and compare with existing assistants.

4. EdotEnv (YC S26) Launches Quant Trading RL Environments to Teach LLMs Research
   T2Hacker News 热门（Algolia · 中文化）

   EdotEnv (YC S26) has launched a suite of reinforcement learning environments for quantitative trading, designed to train large language models in applied ML, long-horizon optimization, and adaptation under adversarial noise. With non-saturating rewards and market-derived data, these environments simulate realistic research scenarios for AI agents. This release aims to lower the barrier to quantitative research and advance LLM-driven financial decision-making and strategy discovery. The environments are specifically built to handle adversarial noise, making them robust for real-world conditions. By integrating these RL environments, researchers and developers can explore novel approaches to trading strategy optimization. EdotEnv's innovation is particularly timely as interest grows in using LLMs for complex financial tasks.

5. MacPaw partners with Liquid AI to offer on-device inference for its app store developers
   T2TechCrunch：AI（RSS）

   MacPaw is building a local version of its AI assistant Eney using Liquid AI models, offering on-device inference to developers building for its app store. This lets developers run AI models directly on user devices without relying on cloud APIs, reducing latency and improving data privacy. The move aligns with the growing trend of on-device AI and could make MacPaw's ecosystem more attractive to AI-native developers.

6. LLM CLI Tool llm 0.32 Released
   T2Simon Willison 博客（Atom）

   Simon Willison released llm 0.32, an open-source tool for accessing various large language models from the command line. The new version likely brings performance improvements, new model support, and plugin updates, helping developers call AI capabilities efficiently in terminal environments and simplifying scripts and automation. It supports multiple mainstream models, making it a practical choice for AI developers.

7. Wrinkles: An AI App That Uncovers the Hidden Stories of the Places Around You
   T2TechCrunch：AI（RSS）

   Wrinkles is an AI-powered audio tour guide available on both iOS and Android that reveals the hidden history and local stories of places around you. Combining location awareness with AI-driven narrative, the app transforms everyday environments into an explorable story map, offering travelers, urban walkers, and history buffs a fresh way to experience the world.

1. Anthropic Appoints Tino Cuellar as Chief Global Affairs Officer
   T1Anthropic：News（网页）

   Anthropic has announced that Tino Cuellar will join the company as Chief Global Affairs Officer. Cuellar, a former Stanford Law School professor, previously served in the U.S. Treasury Department on international affairs and as the U.S. Executive Director at the World Bank. He will lead Anthropic's global policy and government relations efforts, engaging with governments, international organizations, and research institutions on AI regulation, national security, trade rules, and international cooperation. The appointment signals Anthropic's push to elevate its role in global AI governance and policy debates. This move also reflects the company's ambition to shape AI policy at the highest levels. As AI regulation accelerates worldwide, Cuellar's experience in both academia and government will be crucial for Anthropic.

2. OpenAI Clarifies Third-Party Cyber Evaluation Incidents and Announces New Safeguards
   T1OpenAI：News（RSS）

   OpenAI has issued a statement explaining recent third-party cybersecurity evaluation incidents involving its AI models. The company outlines what happened, why external red teams and safety researchers are important, and what new safeguards are being introduced to strengthen AI model testing and evaluation. This includes tighter oversight of third-party evaluators, better verification of evaluation results, and additional risk controls to ensure that safety testing remains transparent and secure. The announcement comes as advanced AI models increasingly attract both security research and potential abuse. OpenAI's move signals a more structured approach to managing external evaluations, which is expected to influence how AI labs conduct vulnerability discovery and disclosure. For the broader AI ecosystem, it highlights the growing importance of standardized evaluation frameworks and coordinated handling of cyber risks. The news matters for AI safety researchers, model rollout teams, and anyone involved in developing or deploying frontier AI systems.

3. Nvidia-Led Open Secure AI Alliance Proposes AI Agent Defenses Just a Week After Launch
   T2TechCrunch：AI（RSS）

   The Open Secure AI Alliance, spearheaded by Nvidia and founded just a week ago, has already grown to over 120 companies and put forward initial proposals for defending against AI agents. The alliance aims to address the emerging security threats posed by AI agents and promote unified industry security standards. Nvidia's swift move signals its push to dominate AI security and cement its leadership in the AI ecosystem, while offering enterprises and developers a more reliable protection framework.

4. TechCrunch Disrupt 2026's Real World AI Stage Features Robots, Automated Factories, and Extinct Animals
   T2TechCrunch：AI（RSS）

   TechCrunch Disrupt 2026 has announced a new Real World AI stage, focusing on the intersection between the digital and physical worlds. The stage will showcase robots applied in real-world scenarios, automated factory implementations, and cutting-edge explorations into bringing extinct animals back with AI. The organizers note that we will continue to see increasing blending of virtual and physical realms, and this stage aims to provide a cross-disciplinary platform for industry players.

1. HalluTruthQA-4K: A Fine-Grained Corpus and Annotation Process for Arabic Hallucination Detection and Truth Verification
   T2arXiv AI/CL/LG（论文）

   Large language models can generate fluent Arabic answers while introducing factual errors that are difficult to identify and verify. Existing Arabic hallucination resources often assign a binary label to an entire response, providing limited information about the exact erroneous content, the reason for the error, or the correct factual answer. This paper presents HalluTruthQA-4K, an expanded version of the HalluTruthQA resource containing 4,000 expert-curated Arabic question-answering instances across four knowledge-intensive domains: Islamic knowledge, history, science, and geography. Each instance pairs an Arabic question with a model-generated response, a verified reference answer, and five plausible distractors. Hallucinated responses are additionally annotated with character-level erroneous spans, human-written explanations, and hierarchical hallucination types. The corpus contains 1,643 hallucinated and 2,357 non-hallucinated responses, with 1,843 annotated erroneous spans. The paper details resource construction, annotation guidelines, taxonomy, data format, inter-annotator agreement, and corpus statistics, providing a reusable resource for hallucination detection, span-level error localization, explanation generation, factual verification, and broader evaluation of factual reliability in Arabic language models.

2. When Efficiency Becomes Fragility: Exploiting Dynamic Routing Vulnerabilities in Adaptive UAV Tracking
   T2arXiv AI/CL/LG（论文）

   This paper reveals that adaptive Transformer trackers, which use input-dependent dynamic routing to balance accuracy and efficiency, harbor a critical structural flaw: the Lipschitz singularity at layer-skipping decision boundaries makes the local Lipschitz constant unbounded. Tiny input perturbations can be amplified at gating modules, causing dramatic changes in the inference topology. The authors formally characterize this singularity in adaptive tracking architectures and, for the first time, identify it as a directly exploitable new attack surface. They propose the Adversarial Path-Inversion (API) framework, which generates imperceptible perturbations to precisely manipulate gating decisions and force the inference onto altered computational paths. The severe inconsistency between original and inverted paths dismantles the model's representation capability. Extensive experiments on state-of-the-art adaptive trackers demonstrate that API achieves superior perturbation stealthiness, more effective attacks, and faster inference speeds. This work opens a new dimension for security analysis of dynamic tracking networks and provides a theoretical warning for constructing robust adaptive tracking architectures in the future.

3. PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
   T2arXiv AI/CL/LG（论文）

   Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents retain preferences, task histories, tool routines, and learned skills across sessions, making them an ideal setting for studying this capability—yet whether retained experience truly improves agents over time has not been systematically tested. PAST-Bench isolates this question through 26 scenarios and 204 episodes, comparing agents with experience retention on and off across memory, procedural reuse, information gathering, and update capabilities. Across seven base models and four agent frameworks, improvement is real but uneven. Guided by these findings, Hermes+ introduces five targeted interventions and raises average gains, strongest on tasks requiring outdated state to be replaced, though effects remain capability- and model-dependent.

4. Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility
   T2arXiv AI/CL/LG（论文）

   Large language models can solve much harder reasoning problems with additional inference-time compute, yet the term 'test-time scaling' now encompasses diverse algorithms that differ in statistical structure, compute accounting, and failure modes. This paper develops a systematic account along three axes: it formalizes test-time scaling as budgeted inference over an autoregressive model's implicit prefix tree, distinguishing single-trajectory, leaf-level, and prefix-level regimes; it treats the whole inference system as the evaluated object, introducing an evaluation profile that recovers or bounds common repeated-sampling metrics and calls for protocol-matched reporting; and it specifies reproducibility requirements separating exact replay from distributional reproducibility. The authors also organize the open-weight reasoning ecosystem and assemble over 2 billion full reasoning traces for release with verifier and token-level signals.

5. Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation
   T2arXiv AI/CL/LG（论文）

   This paper isolates tokenization from all other variables by fixing pretrained Qwen3.5 backbones (0.8B–27B), data, budget, and decoding, then swapping only the music representation across seven tokenizations. The results are striking: representation, not model size, is the binding variable for distributional fidelity. Scaling the backbone 34x barely moves Fréchet Music Distance (FMD), while changing representation halves it. PMT, a performance-resolution stream with 10 ms timing, per-note velocity, multi-track texture, and 609 symbols, reaches FMD 159 at 0.8B, beating beat-grid tokenizations at 27B. The effect persists on a from-scratch 26M model and a second tokenizer, indicating it is a property of the class. Caption adherence is weak but separable; a lightweight decode-time constraint sharply improves instrument and key accuracy. The authors release a harness, 25+ checkpoints, two corpora (86.6k aligned; 6.25M captioned), and a diagnostic showing published systems largely ignore captions.

6. Omega-S: A Functional Resilience Index for LLM Fine-Tuning
   T2arXiv AI/CL/LG（论文）

   This paper presents Omega-S, a regularization penalty computed solely from the weight matrix to mitigate catastrophic forgetting during LLM fine-tuning. It requires no previous-task data, Fisher matrix, or stored old weights, adds only three lines to an existing training loop, and increases per-step cost by under 4%. In LoRA fine-tuning experiments on Llama-3-8B shifting from code to prose, HumanEval pass@1 improved from 0.173 to 0.238, and retention ratio rose from 62.9% to 84.1%, outperforming tuned weight decay and EWC. Mechanistic analysis reveals the actual driver is the node-degree variance term, not the topological eigenvalue term.

7. BanglaWild: An In-the-Wild Bengali Scene Text Recognition Benchmark for OCR and Vision-Language Models
   T2arXiv AI/CL/LG（论文）

   In-the-wild Bengali scene text recognition remains largely unmeasured: existing resources target handwritten documents or constrained sign-board parsing, report only aggregate edit-distance metrics, and evaluate either conventional OCR or VLMs, never both on the same data. This paper introduces BANGLAWILD, a benchmark of 2,535 Bengali scene text images with verbatim gold transcriptions, categorical axes, diagnostic attributes, and orthographic standard forms. Evaluating 15 VLMs and 3 conventional OCR systems under multiple prompting strategies, LoRA fine-tuning, and LLM-as-a-Judge evaluation, the study reveals a persistent gap where larger models do not outperform smaller ones in the same family. Visual mis-recognition accounts for ~60% of errors, while conjunct-related errors contribute under 2%, challenging a long-standing assumption in Bengali OCR research.

8. WorldCup Arena: Prospective, Leakage-Free Evaluation of Frontier LLMs on a Live Tournament
   T2arXiv AI/CL/LG（论文）

   Most benchmarks measuring forecasting ability of large language models are retrospective: the event has already happened, so evaluations must defend against memorisation. This paper flips the design. During the 39 days of the 2026 FIFA World Cup, six frontier LLMs—all with extended thinking and native server-side web search—were asked before each kickoff to fill a seven-market prediction card for all 104 matches, 12 group winners and an outright pool. Since no answer existed when asked, the evaluation is leakage-free by construction. With 4,494 scored predictions, the systems average 63.9% on match outcomes, matching the strategy of backing the bookmaker's favourite. They agree with one another far more often than they are right, so majority vote adds nothing; they under-commit to draws and goals, and crowd scoreline picks onto a single prototypical result. The benchmark, dossiers and scoring code are released.

—