compounding bias
Small mistakes made early in a process can snowball into large errors later, especially when feedback is rare. This is compounding bias: reasoning or generation steps that are slightly off build on each other, leading to major failures before any correction is possible.
The papers highlight how this plays out in different settings. One shows that long reasoning chains accumulate small deviations until weak signals expose them. Another finds that language models' accuracy drops sharply as tasks get longer, as errors pile up. Compression methods like pruning can worsen the effect, increasing bias between groups. Across these, the message is clear: even tiny initial errors compound into significant problems over extended procedures.
- Papers
- 6
- Released code
- 1
- First seen
- Sept 2026
- Latest
- Oct 2026
6 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1Structural guidance improves language models’ long-horizon reasoning under sparse rewards
SAGE improves sparse-reward, long-horizon LLM reasoning by pruning structurally unproductive branches and supplying dense geometric guidance across reasoning depth.
Sept 2026 - 2Long context windows do not ensure reliable sustained execution
Long-Transduction reveals that models suffer 62.8% accuracy loss when context length scales to 128K tokens in long-horizon agentic workflows.
Oct 2026 - 3Compression Widens Demographic Fairness Gaps in Whisper Speech Recognition Models
Post-training pruning and quantization can sharply amplify demographic error burdens in Whisper, while distillation usually narrows them.
Oct 2026
How fast it is moving
Often studied with
agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real1 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.1 togethercredit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.1 togetherThe papers
Most central to this idea first, not most recent.
- Top Universitycs.AIcode
Structural guidance improves language models’ long-horizon reasoning under sparse rewards
SAGE improves sparse-reward, long-horizon LLM reasoning by pruning structurally unproductive branches and supplying dense geometric guidance across reasoning depth.
Virginia Tech, University of Wisconsin Madison · Sept 2026
- Academiccs.CL
Compression Widens Demographic Fairness Gaps in Whisper Speech Recognition Models
Post-training pruning and quantization can sharply amplify demographic error burdens in Whisper, while distillation usually narrows them.
The Ohio State University, Air Force Research Laboratory · Oct 2026
- Chinese Techcs.LG
Execution graph helps off-the-shelf teachers supervise student agents better
Graph-conditioned retrieval of teacher execution histories improves on-policy distillation for compact language agents in multi-turn tasks.
Yuanbao Team, Tencent, Tsinghua University · Sept 2026
- Big Techcs.AI
Long context windows do not ensure reliable sustained execution
Long-Transduction reveals that models suffer 62.8% accuracy loss when context length scales to 128K tokens in long-horizon agentic workflows.
NVIDIA · Oct 2026
- Top Universitycs.RO
Human-like scripted demonstrations reproduce behavioral cloning’s puzzling failures
Behavioral cloning performance can improve with overfitting, and fully closed-loop policies often fail without action chunking.
University of California, Berkeley · Oct 2026
- Big Techcs.CV
Loop closure in streaming 3D reconstruction achieves kilometer-scale accuracy without drift
CLoSeR adds loop closure to a streaming reconstruction backbone to achieve drift-free kilometer-scale 3D reconstruction from monocular video.
ETH Zurich, University of Stuttgart · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.