LongBench is the first bilingual multi-task benchmark for long context understanding in LLMs, containing 21 datasets in 6 categories with average lengths of 6711 words (English) and 13386 characters (Chinese).
hub Mixed citations
Extending Context Window of Large Language Models via Positional Interpolation
Mixed citation behavior. Most common role is background (65%).
74
Pith papers citing it
Background
65% of classified citations
abstract
We present Position Interpolation (PI) that extends the context window sizes of RoPE-based pretrained LLMs such as LLaMA models to up to 32768 with minimal fine-tuning (within 1000 steps), while demonstrating strong empirical results on various tasks that require long context, including passkey retrieval, language modeling, and long document summarization from LLaMA 7B to 65B. Meanwhile, the extended model by Position Interpolation preserve quality relatively well on tasks within its original context window. To achieve this goal, Position Interpolation linearly down-scales the input position indices to match the original context window size, rather than extrapolating beyond the trained context length which may lead to catastrophically high attention scores that completely ruin the self-attention mechanism. Our theoretical study shows that the upper bound of interpolation is at least $\sim 600 \times$ smaller than that of extrapolation, further demonstrating its stability. Models extended via Position Interpolation retain its original architecture and can reuse most pre-existing optimization and infrastructure.
hub tools
citation-role summary
background 14
method 4
baseline 2
citation-polarity summary
representative citing papers
Maven is an RL method using answer-conditioned evidence-state values to assign rewards to add, link, and drop actions on evidence memory, outperforming outcome-only baselines on LongBench v2, LongReason, and RULER.
H2HMem is a multimodal memory benchmark evaluating LLM agents on recall, reasoning, and application in dyadic and multi-party human-human conversations with phenomena such as anaphora and deixis.
LongLive-RAG formulates long video generation as retrieval-augmented generation by treating self-generated latents as a dynamic searchable history and adding a Window Temporal Delta Loss for better retrieval.
Autoregressive transformers exhibit measurable cognitive fatigue during extended generation, quantified by the Fatigue Index that predicts degradation (AUROC 0.95) and repetition (rho 0.94).
Persona Attack uses step-by-step memory injections to achieve up to 95% success in making LLMs ignore safety alignments, with effectiveness depending on model memory and instruction combinations.
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
MemLens benchmark shows long-context LVLMs lose accuracy with length while memory agents lose visual fidelity, with multi-session reasoning below 30% for most systems and neither approach solving the task alone.
ExtraVAR enables resolution extrapolation in visual autoregressive models by stage-aware RoPE remapping and entropy-driven attention scaling, suppressing repetition and detail loss.
Jordan-RoPE realizes a distance-modulated phase basis via non-semisimple Jordan blocks, generating features such as d e^{iωd} for relative positional encoding.
Dual Triangle Attention achieves effective bidirectional attention with built-in positional inductive bias via dual triangular masks, outperforming standard bidirectional attention on position-sensitive tasks and showing strong masked language modeling results with or without positional embeddings.
Multiscreen replaces softmax attention with screening to provide absolute query-key relevance, resulting in models with 30% fewer parameters that maintain stable performance at long contexts.
SHARP applies a spectrum-aware dynamic RoPE scaling schedule that promotes resolution more strongly in early denoising stages and relaxes it later, outperforming static baselines on quality metrics for remote sensing images.
GRAPE unifies RoPE and ALiBi as special cases of group actions on positions, providing a principled design space for positional encodings via SO(d) rotations and GL unipotent transformations.
Stacked causal self-attention combined with LayerNorm induces recency bias in Transformer decoders, reversing the earlier-token bias seen in attention alone.
KV cache compression causes task-dependent degradation in high-density reasoning due to disrupted CoT links; ShotKV mitigates this by preserving few-shot examples as indivisible semantic units through phase separation, delivering 9-18% accuracy gains and 11% latency reduction.
Infini-attention combines compressive memory with masked local attention and long-term linear attention inside each Transformer block to support infinite context length with bounded resources.
LongRoPE extends LLM context windows to 2048k tokens via search for non-uniform positional interpolation, progressive fine-tuning from 256k, and short-context readjustment.
ACE is a pluggable module that elastically orchestrates historical agent steps as raw, abstract, or dropped to maintain compact yet recoverable context for LLM agents handling long trajectories.
LPES uses per-layer scaling factors optimized by a genetic algorithm with Bézier curves to balance attention and improve long-context LLM performance by up to 11.2% on key-value retrieval.
RoVE rotates value embeddings simultaneously with keys in attention to make values position-dependent, reframing RoPE as attentive convolution and reporting gains on long-context tasks in 124M and 354M GPT-2 models.
IS-CoT framework interleaves planning, writing, and reflection in LLMs to prevent length collapse, yielding IS-Writer-8B that outperforms larger models on long-form benchmarks with better length compliance.
PJ-RoPE organizes relative-position mechanisms as a learnable Fourier-Jet-Affine space derived from lag-shift dynamics, extending RoPE and ALiBi with explicit jets and sector selection.
Language models show superior memory to humans on psych experiments but can be adjusted via prompting and compaction to forget more human-like, yielding better user simulators.
citing papers explorer
-
LayerNorm Induces Recency Bias in Transformer Decoders
Stacked causal self-attention combined with LayerNorm induces recency bias in Transformer decoders, reversing the earlier-token bias seen in attention alone.
-
Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression
KV cache compression causes task-dependent degradation in high-density reasoning due to disrupted CoT links; ShotKV mitigates this by preserving few-shot examples as indivisible semantic units through phase separation, delivering 9-18% accuracy gains and 11% latency reduction.
-
MTraining: Distributed Dynamic Sparse Attention for Efficient Ultra-Long Context Training
MTraining scales LLM training to 512K-token contexts on 32 A100 GPUs by integrating dynamic sparse training patterns with balanced and hierarchical sparse ring attention, achieving up to 6x throughput gains without accuracy loss on long-context benchmarks.
-
Positional Encoding via Token-Aware Phase Attention
TAPA adds a learnable phase function to attention to preserve long-range token interactions, enabling direct continual pretraining, length extrapolation, lower perplexity, and stronger retrieval than RoPE-style methods.
-
MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent
MemAgent uses multi-conversation RL to train a memory agent that reads text in segments and overwrites memory, extrapolating from 8K training to 3.5M token QA with under 5% loss and 95%+ on 512K RULER.
-
Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
Applying a head-specific sigmoid gate after SDPA in LLMs boosts performance and stability by adding non-linearity and query-dependent sparse modulation while reducing attention sinks.
-
A Survey of Context Engineering for Large Language Models
The survey organizes Context Engineering into retrieval, processing, management, and integrated systems like RAG and multi-agent setups while identifying an asymmetry where LLMs handle complex inputs well but struggle with equally sophisticated long outputs.
-
Gemma 3 Technical Report
Gemma 3 introduces multimodal open models with architectural changes for efficient long context, trained via distillation and a new post-training recipe that makes the 4B version competitive with prior 27B models and the 27B version comparable to Gemini-1.5-Pro.