The Archive

RL-based approach teaches LLMs tool-use patterns for calculator integration on Countdown task, with supervised pre-training and on-policy refinement.

Empirical study shows sliding-window attention outperforms linear attention variants on efficiency-performance tradeoff for LLMs.

PLVR learns verifiable symbolic programs for LLM reasoning instead of embedding capability in weights, enabling interpretability and transfer.

Study showing Vision Transformer attention heads in multimodal LLMs specialize into object/background roles; SHS-Index metric correlates specialization with downstream performance.

Study across 30 LLMs showing linguistic confidence diverges from internal confidence (logits/semantic entropy) on classification and generation tasks.

MAIL framework applies memory-augmented LLMs to hypothesis generation in chemistry by adaptively navigating chemical literature.

Real-time semantic navigation framework for UAV object-goal search using LLMs to reason over spatial context and prioritize regions.

MCR-Bench benchmarks LLMs on iterative, multi-round code review tasks capturing real-world defect-state interactions.

Linear probe analysis of how open-weight LLMs organize Moral Foundations Theory categories in representation space.

CorporateBench: 230k-document Q&A benchmark for evaluating LLMs on enterprise document retrieval and knowledge base queries.

LLMs can synthesize near-optimal algorithms for operations research problems (inventory control, queueing, assortment) given problem class descriptions and parameter ranges.

STEP: Multi-modal LLM system for human-robot collaboration that tracks state transitions to reduce hallucinated actions in task planning.

BALMS: Benchmark for evaluating agentic LLMs on longitudinal mental health inference from wearable signals.

A recap of all the incidents involving LLMs made by Anthropic, Meta, and OpenAI, which went rogue and attacked real companies and individuals on the internet.

Framework for continual post-training of multimodal LLMs on unlabeled data, using token-level visual dependence to prevent catastrophic forgetting.

AsymSpec: asymmetric speculative decoding for agentic LLMs that compresses verifier context while drafting on full input to reduce inference latency.

Layer-wise MBTI personality analysis of quantized LLMs reveals trait changes across precision levels.

SciMIF benchmark evaluates MLLMs' complex instruction-following across 22 tasks in 5 scientific domains.

Study shows magnitude pruning degrades sparse autoencoders' interpretability of LLMs; activation-aware methods like Wanda preserve SAE robustness.

Code World Model combines LLMs and video models to separate world dynamics from visual rendering for coherent, open-ended environment simulation.

DEDUCE framework enables LLMs to detect and correct factual errors in user inputs rather than passively accepting misleading premises.

Multilingual self-play evaluation reveals LLMs exhibit significant cross-lingual skill inconsistency independent of knowledge gaps.

The startup came out of stealth with $6 million in seed funding and a plan to use LLMs and cybersecurity know-how to make AI queries coherent.

The focus is on agentic capability and predictable enterprise deployment.

LLMs exhibit poorly calibrated confidence in hidden-information tasks, with high-confidence errors in agentic decision-making contexts.

BPCO training recipe stabilizes critic-based RL for LLMs by combining DPPO, bounded value predictions, and adaptive advantage estimation.

SRPO framework enables LLMs to self-reflect on trajectories and synthesize errors into dense training signals for improved long-horizon reasoning.

StrategyBench evaluates whether LLMs can induce explicit task rules from examples in few-shot learning, addressing sensitivity to example construction.

ProxyFormer reduces quadratic attention overhead in ultra-long-context LLMs via dual-stream proxy token architecture with bottom-up compression and top-down decompression.