AuthorsZecheng Tang, Baibei Ji, Ruoxi Sun et al.
TL;DR
MemoryRewardBench uses long-context reasoning, multi-turn dialogue, and long-form generation settings to reveal that GLM4.5-106A12B reaches 68.21 average accuracy and narrows the gap with proprietary reward models.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory benchmarks evaluate LLMs directly, while MemoryRewardBench highlights that automated RM-based memory assessment remains largely unexplored and unreliable.
When reward models misjudge intermediate memories in long-context reasoning or multi-turn dialogue, downstream supervision fails, degrading long-term consistency, temporal reasoning, and constraint-following.
HOW IT WORKS
MemoryRewardBench builds on Sequential Pattern, Parallelism Pattern, Mixed Pattern, Long-context Reasoning, Multi-turn Dialogue Understanding, and Long-form Generation to generate paired memory trajectories.
Think of MemoryRewardBench as a stress-test lab where different memory pipelines are like distinct circuits, and reward models act as judges scoring how cleanly information flows.
This design lets MemoryRewardBench probe process-based memory quality, not just final answers, revealing where plain long context windows fail to capture nuanced memory management.
DIAGRAM
This diagram shows how MemoryRewardBench instantiates long-context reasoning, multi-turn dialogue understanding, and long-form generation tasks with specific memory patterns.
DIAGRAM
This diagram shows how MemoryRewardBench feeds paired trajectories to reward models and computes judgment accuracy across 10 settings and 2,400 examples.
PROCESS
Memory Management Pattern
MemoryRewardBench first selects Sequential Pattern, Parallelism Pattern, or Mixed Pattern to define how intermediate memories evolve over chunks.
Task Overview
MemoryRewardBench instantiates Long-context Reasoning, Multi-turn Dialogue Understanding, or Long-form Generation with context lengths from 8K to 128K tokens.
Benchmark Construction
MemoryRewardBench builds chosen and rejected trajectories using NOISE, DROP, Mem0, AMem, and constraint perturbations while keeping outcomes or processes controlled.
RM Evaluation Criteria
MemoryRewardBench applies outcome-based and process-based criteria, asking reward models to pick the better trajectory and justify their choice.
KEY CONTRIBUTIONS
Introduce MemoryRewardBench
MemoryRewardBench is the first benchmark targeting reward models, with 10 settings and 2,400 examples spanning 8K to 128K-token contexts for memory management evaluation.
Define memory management patterns
MemoryRewardBench formalizes Sequential Pattern, Parallelism Pattern, and Mixed Pattern, showing how intermediate memories M evolve and fuse to produce outcomes.
Cross-task RM analysis
MemoryRewardBench compares 13 reward models on long-context reasoning, multi-turn dialogue, and long-form generation, revealing GLM4.5-106A12B at 68.21 vs Claude-Opus-4.5 at 74.75 average accuracy.
RESULTS
Avg.
74.75
+6.54 over GLM4.5-106A12B
Long-context Reasoning Avg.
68.88
Claude-Opus-4.5 vs 69.13 for GLM4.5-106A12B
Multi-turn Dialogue Avg.
64.07
Claude-Opus-4.5 vs 56.75 for GLM4.5-106A12B
Long-form Generation Avg.
74.75
Claude-Opus-4.5 vs 68.21 for GLM4.5-106A12B
MemoryRewardBench evaluates reward models on 10 settings built from LongBench, RULER, LoCoMo, LongEval, LongGenBench, and LongProc. The 74.75 average accuracy for Claude-Opus-4.5 shows that current best reward models still make roughly one in four incorrect memory management judgments.
BENCHMARK
MemoryRewardBench evaluates reward models on 10 settings built from LongBench, RULER, LoCoMo, LongEval, LongGenBench, and LongProc. The 74.75 average accuracy for Claude-Opus-4.5 shows that current best reward models still make roughly one in four incorrect memory management judgments.
BENCHMARK
Average accuracy on MemoryRewardBench over all task settings.
KEY INSIGHT
MemoryRewardBench shows Qwen3-4B at 52.42 average accuracy, surpassing the larger Qwen2.5-7B which only reaches 38.17.
This breaks the common assumption that more parameters always help reward modeling, highlighting a strong generational advantage over raw model size.
WHY IT MATTERS
MemoryRewardBench gives practitioners a concrete way to benchmark reward models on long-term memory management rather than only final-answer quality.
Builders can now choose and train reward models that explicitly favor clean, coherent memory trajectories, enabling more reliable supervision of memory-centric LLM systems.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…