Skip to main content
Search

Baidu (China)

5 related articles

Subscribe to updates
  1. arXiv

    GRPO safety alignment loses its signal on hard prompts: Beacon restores a safe direction and cuts average safety error by up to 32.83 points across four models

    Safety alignment of large reasoning models commonly relies on GRPO, but responses sampled for one prompt often share the same safety outcome, so the within-group relative advantage vanishes and the model gets no direction toward safety on persistently unsafe prompts. Beacon pairs safety signal recovery, which filters high-contrast trajectories and directly supervises a safe reference response when the signal collapses, with safety-utility decoupled alignment, which assigns category-aware rewards and applies KL anchoring only to benign prompts. Across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks, average safety error falls by 17.69 to 32.83 percentage points relative to the initial policies, with XSTest over-refusal between 0.40% and 1.20%.

  2. arXiv

    Agents must ask real clients what they left unsaid: across 49 projects, the best delivery score is 0.5510

    FDE-Bench turns 49 real customer projects into end-to-end delivery tasks where agents must elicit unstated requirements from a simulated customer and iterate, and across 18 agent configurations the highest mean delivery score is only 0.5510 against an FDE reference of 1.0.

  3. arXiv

    MIRROR turns LLM personalization from imitation to preference internalization via reference-revealed on-policy self-distillation, leading three benchmarks with less forgetting

    MIRROR treats a user's reference text as hindsight information and aligns the model's next-token distributions along its own on-policy trajectories with those of its reference-conditioned self, internalizing preferences instead of copying wording; MIRROR-F adds selective anchoring on informative, context-grounded reference tokens. Across three personalized generation benchmarks, two model scales, and reference-based plus LLM-judge evaluations, both achieve leading personalization and text quality with less forgetting than SFT baselines on three unseen tasks.

  4. arXiv

    FactorEngram factorizes n-gram memory over a shared dictionary, lifting 8K long-context retrieval by up to 43.3 points at 340M and 1B backbones

    The work proposes FactorEngram, which replaces monolithic per-pattern n-gram embeddings with sparsity-regularized coefficients over a dictionary shared across patterns and reuses that same dictionary for basis-level contextual gating, improving language modeling, downstream tasks, and long-context retrieval on 340M- and 1B-parameter Transformer backbones.

  5. USTC and Baidu team survey parallel reasoning: a unified formal framework maps non-interactive, interactive, and efficiency methods into one roadmap

    This survey defines parallel reasoning as a three-stage inference paradigm of decomposition, parallel processing, and aggregation, gives a formal formulation and distinguishes it from Chain-of-Thought and long-thinking, then organizes representative methods along three lines—non-interactive (self-consistency, Best-of-N ranking, structured reasoning), interactive (intra-model and multi-agent interaction), and efficiency (parallel decoding, parallel function calling, speculative decoding)—and summarizes applications, core challenges, and future directions.

You’re all caught up