Pulkit Aneja

Initially, the experience was positive, but it later became frustrating as the system started providing incorrect and repetitive answers. Upon reflection, I identified a few reasons for this issue: - The current retrieval system is too coarse, matching queries to the closest “question/topic” but struggling with variations, such as “symptoms of diabetes.” - The performance also depends on the dataset size; I uploaded relatively small files for my understanding. - TF-IDF sometimes selects the wrong match, like nutrition tips instead of FAQ answers. This leads to the perception that the agent is stuck providing irrelevant responses. To enhance robustness, consider the following strategies: - Expand the documentation. - Chunk large documents for better processing. - Improve the retrieval system. - Implement fallback rules to prevent nonsensical answers. To be continued… I’ll retrain my “nutrition‑obsessed” agent into something more balanced. Moral of the story: even AI agents need a diversified diet. 😅 learning starts and experience gains only when mistake happens.

8

Foundation models have been extensively utilized in time series forecasting, with models like TimeGPT and TimesFM gaining significant attention. 𝗖𝗵𝗿𝗼𝗻𝗼𝘀 is a forecasting model developed by Amazon, that is based on large language model architecture. Chronos models have been trained on a huge corpus of time series datasets, as well as synthetic data generated with Gaussian processes. Chronos-2 has recently been released, achieving the top performance on GIFT-Eval amongst pretrained models! Chronos can be downloaded on Hugging Face, or easily accessed via the Darts Python library. Visit the link below for more information and make sure to follow me for regular data science content! 𝗖𝗵𝗿𝗼𝗻𝗼𝘀 𝗚𝗶𝘁𝗵𝘂𝗯 𝗽𝗮𝗴𝗲: https://lnkd.in/dvQ6D6EA 𝗖𝗵𝗲𝗰𝗸 𝗮 𝗳𝗿𝗲𝗲 𝗠𝗟 𝗯𝗼𝗼𝗸: https://lnkd.in/dyByK4F

304

10 Comments

🦾🧠 Building Deep Research Agents ?? Adding think tags can collapse training making model stuck in long reasoning trajectories Authors revamped search-R1 strategy for deep research. They decoupled the entire deep research dimensions along with enhancements as following: > prompt template: proves fast thinking templates are more stable and generates short response length. Slow thinking templates get stuck in <think> tags > modified F-1 penalizes the reward to discourage omitting search or answer steps. This makes the policy to avoid being collapsed into a simpler state of generating no answer > policy optimization: authors found GRPO gets noisier in multi-step long context reasoning. PPO gets challenging as it tries to fit value function over sparse outcome rewards (Exact Match based). Found REINFORCE stays table with lowest search frequency Papers shows high promises while introducing simpler updates to search-R1. Overall I am intrigued by the findings that slow thinking could: > de-stabilise the training > longer reasoning chains > multiple search queries > often leading to answer avoidance

3

Cov(X, X) ≠ 1. I was working on a proof with a student when he wrote: Cov(X, X) = 1 It felt right. Same variable. Perfect relationship. So… 1? Not quite. That’s not covariance. That’s correlation. Here’s the difference: - Covariance (X, X) = Variance (X) - Correlation (X, X) = 1 Always. That one small mix-up? It broke the entire proof. Small confusion. Big consequences.

48

Wrote a blog explaining "LoRA Fine-tuning Efficiency Under Different Loss Functions." LoRA made fine-tuning cheap, but not always efficient. This post explores how loss functions like Cross-Entropy, Label Smoothing, and Focal Loss change convergence, stability, and generalization in lightweight LLMs. Read here: https://lnkd.in/gh_vm2kG

2

AI Tip: RL for LLMs; Real Gains, Real Risks ⚡🤖 Reinforcement learning (RL) can supercharge your LLMs; but only if you apply it intentionally. 🚀 Where RL Delivers ▫️Stronger alignment & safety: RLHF makes models follow instructions more reliably and reduces harmful outputs. ▫️Domain-specific behaviour: RL-tuned models behave like rule-bound assistants, ideal for regulated industries. ▫️Lower-cost tuning: Newer approaches like PE-RLHF deliver alignment gains with far less compute. ⚠️ What to Watch ▫️Human feedback is expensive, scaling high-quality labels is the bottleneck. ▫️Reward-gaming: Poor reward design = models optimise for the wrong behaviours. ▫️Drift & governance: RL-tuned models need monitoring, evaluation, & version control. ✅ What to Do Next ▫️Start with a small RL pilot in a low-risk workflow. ▫️Use efficient RL variants if compute is limited. ▫️Monitor drift: track error rate, safety incidents, or task-completion rate post-RL. 👇 If you had to track ONE metric after RL fine-tuning, what would it be? 📚 🔗 https://lnkd.in/gtCuJmMR 🔗 https://lnkd.in/gUQebVdN 🔗 https://lnkd.in/gKsxPpN6 🔗 https://lnkd.in/g8HFaRDJ 🔁 LIKE & RE-SHARE if you enjoyed this post 👥 Follow for more AI fundamentals, tips and use-cases #AI #AITip #ReinforcementLearning #RLHF #LLMs #Alignment #AIEthics #MachineLearning

2

How students think ML hiring works: Strong DSA skills Theory-heavy resume Multiple certifications How ML hiring actually works: Deployed projects with measurable impact Clear explanation of model choices using frameworks System design ability for production ML Evidence of learning speed through iteration Make recruiters notice your signal through clarity, not chaos. Don't rely on applications alone when frameworks create better paths. #AIHiring #CareerReality

14

Online A/B Testing Learning Resources A friend of mine asked me what the best public learning resources are for him to learn applied A/B testing in industry when he has no direct experience. I have suggested a few resources with him, which are also shared below. There are quite a few tech blogs on the same topic, which I will share later. (1) Advanced A/B testing workshop: This is from Elea Feit (Marketing Science professor at Drexel). She is an expert in experimentation and has delivered good lectures and tutorials on this topic, which are public.  https://lnkd.in/gfuzj_jT https://lnkd.in/gjAiJ5ck  https://lnkd.in/ggqkg4GA (2) Udacity - Introduction to A/B Testing: Free course from Google experimentation experts (Diane Tang, Carrie Gimes, Caroline Bucky). Good as an intro  https://lnkd.in/gUBtN26D   (3) Trust-worthy online experimentation: This is from Ron Kohavi, Diane Tang and Ya Xu. Considered as the best book on this topic of applied A/B testing. https://lnkd.in/g6zNZfCQ (4) Causal Inference and Its Application in the Industry: Online book from Alex Deng https://lnkd.in/gCgcjtDi #ABTesting #CausalInference #ControlledExperiments

263

3 Comments