LLMs & Reasoning

Pandora simulates psychotherapy clients by separating psychological state and content planning from an individual expression planner, improving expression similarity and reducing perceived AI-like judgments in the reported evaluation.

LoCo learns state-conditioned continue-or-conclude preferences from self-generated math traces by freezing overthinking prefixes and training DPO on shorter verified continuations, reducing reasoning length while preserving or improving accuracy.

Recursa adapts a foundation model system to a new domain without human experts or a stronger teacher model by bootstrapping a knowledge repository of instructions, skills, and memory from domain materials, then refining it through parallel multi-agent recursive self-improvement, raising all-pass rate on Harvey LAB's Firm Knowledge benchmark from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

Order-Only On-Policy Distillation trains a student to match the teacher’s ordering of top-K tokens at student-generated prefixes, avoiding teacher probability values and temperature tuning.

IdeaTaste benchmarks LLM research taste by comparing proposals written from a 2025 literature snapshot with papers human communities published in 2026, finding LLM agents match citations and ideas far below human levels and mainly fail to turn directions into concrete approaches.

FLAMIA is a family of Blackwell attention kernels with MXFP8 and mixed NVFP4/FP8 training paths that use token-block scaling, causal-leakage safeguards, forward-backward numerical consistency, and GPU-side scheduling of packed documents, achieving 1.10–1.40× (B200) and 1.06–1.42× (B300) speedups on combined forward and backward attention calls.

DecisionRank trains a single-pass decision reranker by combining offline RankNet supervision with item-space on-policy distillation from a generative teacher that provides conditional feedback over candidate items.

The paper argues a qualified yes: co-designing a lightweight System-1 predictor with a costly System-2 executor can reduce execution compute toward 1/ρ_u where interactions are costly, unevenly informative, and predictable before execution.

WeFair trains a compact personalized short-video query rewriter by matching human-feedback fidelity to curriculum SFT, consumption-pair DPO, and strict-target GRPO, reducing intent drift and improving online engagement.

The paper proves for JEPAs that separately trained modalities with shared positive pairs learn the same representation up to a rotation, and shows empirically that good views align better and more efficiently than bad views or next-token prediction.

Soft-MTP removes the noisy residual from the standard multi-token prediction gradient by recovering the KL self-distillation gradient exactly from existing cross-entropy computations, improving NTP, speculative decoding, and zero-shot accuracy without added cost.

Glance equips a single resident 9B generator backbone with a lightweight decision adapter and pointer head to rerank candidate passages from one shared prefill and then generate answers with the same backbone.

AR-BED trains a reasoning LLM with information-gain rewards under an externally specified finite Bayesian experimental-design model, improving adaptive questioning and transferring to a clinical dialogue task.

LongZero lets an LLM improve long-context reasoning from zero external data by first writing structured question-evidence-answer specifications, expanding them into long multi-document contexts, adding self-written retrieval conversations, and training with GRPO.

The paper instruments atomic single-token word-replacement rules and a binned logit model to measure instruction capacity and degradation laws, reporting a GPT-4o mini passport with capacities in rules and uncertainties.